Data processing system

Abstract

The present application relates to a kind of customer information management systems, information acquisition unit is automatically collected enterprise name, representative name, financial information, historical events all certificates, personal information protection system, past business performance, administrative guidance record, website content, alliance party, business plan and other information from the diverse databases and external information sources inside and outside the company, determination unit uses generative AI to multidimensional analysis and scoring on these information, comprehensively determines suitability and risk, thereby unifying and high-precision automating customer investigation and review business previously carried out by hand and complicated inefficient, in addition, mapping unit visualizes the determination result and risk assessment according to multiple attributes, thereby realizing governance strengthening, risk reduction, rapid decision-making in the company.

Description

Technical Field

[0001] The technology disclosed herein relates to a system. Background Technology

[0002] Patent Document 1 discloses a personalized chatbot control method executed by at least one processor, the method comprising: receiving user speech; adding the user speech to a prompt containing instructions related to a chatbot role; encoding the prompt; and inputting the encoded prompt into a language model to generate chatbot speech in response to the user speech.

[0003] Patent document 1: Japanese Patent Application Publication No. 2022-180282.

[0004] The problem this invention aims to solve is that multiple departments within a company, such as customer management, legal compliance, accounting and finance, and sales, individually and manually conduct investigations and reviews when signing contracts with new clients or managing existing clients. These reviews involve various aspects, including franchisee investigations, financial investigations, verification of all historical documents, confirmation of personal information processing systems, investigations of past business performance, confirmation of no overlap with existing clients, suitability assessments of affiliates, appropriateness evaluations of business plans, verification of administrative guidance records, and verification of the authenticity of website content. Therefore, obtaining and verifying information requires a significant investment of time and labor. Difficulties in information sharing between departments and the lack of a unified understanding of past review processes have led to problems such as duplication or omission of investigation content, deviations in judgment standards, neglect of risks, compliance violations, reduced business efficiency, and delayed decision-making. The solution lies in fundamentally improving the efficiency of previously cumbersome and inefficient customer investigation and review processes by implementing a comprehensive database containing past, present, and future predictions of customer information; objective and highly accurate suitability assessment based on AI; and visualization of relationships and risks based on enterprise maps. This provides a new foundation for customer information management that enables strengthened corporate governance, reduced risks, and rapid decision-making. Summary of the Invention

[0005] The means to solve the problem involved in this invention is to provide a customer information management system. This system includes an information acquisition department, a judgment department, and a mapping department. The information acquisition department automatically collects multi-dimensional enterprise information from multiple databases inside and outside the company and external information sources, including the company name, representative name, financial information for multiple past periods, all historical certificates, personal information protection system, past business performance, administrative guidance records, website content, alliance information, business plans, and geographical information of potential new customers. The judgment department uses generative AI to analyze the acquired information, considering factors such as financial soundness, legitimacy of legal person status, suitability of personal information management system, reliability of past business performance, and the presence or absence of administrative guidance. The system scores and determines suitability based on multiple factors, including the authenticity of records and website content, the appropriateness of affiliates, the suitability of business plans, and whether there is overlap with existing clients. Based on the assessment results, it automatically outputs whether registration is permissible and reasons for inapplicability. The mapping department categorizes registered client information by attributes such as region, industry, transaction size, and risk level, visualizing it as a corporate map. Past registration history and assessments are managed in a database. Furthermore, the assessment department uses a structure that continuously learns from past assessment results and external information to improve assessment accuracy. This streamlined and automated process of previously cumbersome and inefficient client investigation and review enhances internal governance, reduces risk, and facilitates rapid decision-making. Attached Figure Description

[0006] Figure 1 This is a conceptual diagram illustrating an example of the configuration of a data processing system according to the first embodiment.

[0007] Figure 2 This is a conceptual diagram illustrating an example of the main functions of the data processing apparatus and smart device according to the first embodiment.

[0008] Figure 3 This is a conceptual diagram illustrating an example of the data processing system configuration in the second embodiment.

[0009] Figure 4 This is a conceptual diagram illustrating an example of the main functions of the data processing device and smart glasses according to the second embodiment.

[0010] Figure 5 This is a conceptual diagram illustrating an example of the data processing system configuration in the third embodiment.

[0011] Figure 6 This is a conceptual diagram illustrating an example of the main functions of the data processing device and head-mounted terminal according to the third embodiment.

[0012] Figure 7 This is a conceptual diagram illustrating an example of the data processing system configuration in the fourth embodiment.

[0013] Figure 8 This is a conceptual diagram illustrating an example of the functions of the main parts of the data processing device and robot according to the fourth embodiment.

[0014] Figure 9 It represents an emotion graph that maps multiple emotions.

[0015] Figure 10 It represents an emotion graph that maps multiple emotions.

[0016] Explanation of reference numerals in the attached figures Data processing systems 10, 210, 310, and 410 12 Data processing device 14 Smart devices 214 Smart Glasses 314 Head-mounted terminal 414 Robot. Detailed Implementation

[0017] Hereinafter, an example of an implementation of the system involved in this disclosure will be described with reference to the accompanying drawings.

[0018] First, let's explain the terms used in the following description.

[0019] In the following embodiments, the processor (hereinafter referred to as "processor") can be a single computing device or a combination of multiple computing devices. Furthermore, a processor can be a single computing device or a combination of multiple computing devices. Examples of computing devices include CPU (Central Processing Unit), GPU (Graphics Processing Unit), GPGPU (General-Purpose computing on Graphics Processing Units), APU (Accelerated Processing Unit), etc.

[0020] In the following implementation, the labeled RAM (Random Access Memory) is a memory that temporarily stores information and is used by the processor as working memory.

[0021] In the following embodiments, the labeled memory is one or more non-volatile storage devices used to store various programs and various parameters, etc. Examples of non-volatile storage devices include flash memory (SSD (Solid State Drive)), disk (e.g., hard disk) or magnetic tape, etc.

[0022] In the following implementation, the labeled Communication I / F (Interface) is an interface that includes a communication processor and an antenna, etc. The Communication I / F is responsible for communication between multiple computers. Examples of communication standards applicable to the Communication I / F include wireless communication standards such as 5G (5th Generation Mobile Communication System), Wi-Fi (registered trademark), or Bluetooth (registered trademark).

[0023] In the following implementation, "A and / or B" is synonymous with "at least one of A and B". That is, "A and / or B" means that it can be only A, only B, or a combination of A and B. Furthermore, in this specification, when "and / or" connects more than three items, the same approach as "A and / or B" applies.

[0024] [First Implementation] Figure 1 An example of the configuration of the data processing system 10 according to the first embodiment is shown.

[0025] like Figure 1 As shown, the data processing system 10 includes a data processing device 12 and an intelligent device 14. An example of the data processing device 12 is a server.

[0026] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0027] The smart device 14 includes a computer 36, a receiver 38, an output device 40, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the receiver 38, output device 40, camera 42, and communication I / F 44 are also connected to the bus 52.

[0028] The receiving device 38 includes a touchscreen 38A and a microphone 38B, etc., for receiving user input. The touchscreen 38A receives user input generated by contact with an indicator (e.g., a pen or finger). The microphone 38B receives user input generated by sound by detecting the user's voice. The control unit 46A in the processor 46 sends data representing user input received through the touchscreen 38A and microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the data representing the user input.

[0029] The output device 40 includes a display 40A and a speaker 40B, and presents data to the user 20 by outputting data in a form perceptible to the user 20 (e.g., sound and / or text). The display 40A displays visual information such as text and images according to instructions from the processor 46. The speaker 40B outputs sound according to instructions from the processor 46. The camera 42 is a small digital camera equipped with an optical system such as a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor.

[0030] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for sending and receiving various information between processor 46 and processor 28 via network 54.

[0031] Figure 2 An example of the main functions of the data processing device 12 and the smart device 14 is shown.

[0032] like Figure 2 As shown, in the data processing apparatus 12, specific processing is performed by the processor 28. A specific processing program 56 is stored in the memory 32. The specific processing program 56 is an example of a "program" as understood in this disclosure. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 based on the specific processing program 56 executed on the RAM 30.

[0033] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0034] In the smart device 14, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The receive / output program 60 is used in conjunction with the data processing system 10 and the specific processing program 56. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart device 14 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290. The receive / output processing is implemented by the processor 46 acting as a control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0035] In addition, other devices besides the data processing device 12 may also have a data generation model 58.

[0036] For example, a server device (e.g., a generation server) may have a data generation model 58.

[0037] In this case, the data processing device 12 obtains the processing results (prediction results, etc.) using the data generation model 58 by communicating with a server device that has the data generation model 58. Furthermore, the data processing device 12 can be a server device or a user-owned terminal device (e.g., a mobile phone, robot, home appliance, etc.). Next, an example of the processing performed by the data processing system 10 of the first embodiment will be described.

[0038] (Example 1) The specific processing flow in Example 1 will be described below. Each part of the system described below is implemented by the data processing device 12 and the smart device 14. Furthermore, the data processing device 12 is referred to as the "server" and the smart device 14 is referred to as the "terminal".

[0039] (Implementation methods for carrying out the present invention) As an embodiment of the present invention, the customer information management system is constructed by connecting a server and a terminal via a network. Multiple functional modules are installed on the server side, including an information acquisition unit, a judgment unit, a mapping unit, a database management unit, an external linkage unit, an AI learning management unit, a log management unit, and a security management unit. On the terminal side, a user interface unit, an input support unit, a result display unit, a notification unit, a file upload unit, a download unit, and a user authentication unit are installed. Specifically, this customer information management system is built in a distributed computing environment with multiple computing processing devices and large-capacity storage. The system containerizes each functional module on the server side as an independent microservice. The system uses container orchestration technology to dynamically allocate execution resources for each module. The computing processing devices on the server side include a central computing processing unit and an image processing unit dedicated to parallel computing. The system utilizes these hardware resources to perform large-scale data processing at high speed. The information acquisition unit performs asynchronous communication via a network interface. The judgment unit drives large-scale language models and deep learning models deployed in memory. The mapping unit executes dimensionality reduction algorithms that project high-dimensional data into a low-dimensional space. The database management unit comprehensively controls relational databases and distributed non-relational databases. The External Linkage Department communicates with external application programming interfaces (APIs) using secure communication protocols. The AI ​​Learning Management Department uses the parallel computing core of the image processing unit to update the weight parameters of the neural network. The Log Management Department writes system events to a time-series database at high speed. The Security Management Department uses a hardware encryption module to encrypt and decrypt data. The terminal side has a computing device capable of running web browsers or dedicated applications. The User Interface Department dynamically rewrites the Document Object Model to depict the screen. The Input Support Department performs predictive transformations by referring to dictionary data cached in the local memory on the terminal side. The Results Display Department parses the structured data received from the server and renders graphical elements. The Notification Department uses WebSocket communication to receive push notifications from the server in real time. The File Upload Department divides files into data blocks of a certain size and sends them asynchronously. The Download Department writes the received data stream to local storage. The User Authentication Department performs challenge-response authentication using public-key cryptography. Through these structures, the system appropriately distributes the processing load across the server and terminals. This system improves the overall throughput of the system. This system is not simply business automation, but rather brings about a technological improvement in processing speed through optimization of the computer architecture. Specific application areas of this system include credit review systems in financial institutions and supply chain risk management systems in global enterprises.

[0040] First, end-users (e.g., heads of customer management, legal, accounting, sales, and audit departments) input or upload necessary information for new customer candidates in various formats through the user interface. This information includes the company name, representative name, location, contact information, establishment date, capital, number of employees, business scope, financial statements for the past three periods (e.g., balance sheet, profit and loss statement, cash flow statement), all historical documents (e.g., PDF, image data), information on personal information protection mechanisms (e.g., privacy policy, information management regulations, office access control system, implementation status of information security training), past business performance (e.g., list of major clients, past administrative guidance records, past representative experience, past disputes with clients), administrative guidance records (e.g., details of administrative penalties, date of penalty, improvement report), company website URL, alliance information (e.g., partner company name, cooperation details, cooperation start date), business plan (e.g., sales targets, investment plan, growth strategy, risk analysis), and geographic information (e.g., prefecture, city / town / village, business office location). For example, financial statements are uploaded as Excel files or PDFs, and all historical certificates are attached as image files. Regarding personal information protection mechanisms, data is entered in the form of checklists or document files; information on past business performance can also be uniformly registered via text input or CSV files. Furthermore, external information can be automatically retrieved by entering the company website URL. Specifically, the terminal-side processing unit performs preprocessing to convert the diverse input data formats into a unified internal data structure. This terminal uses a character code normalization algorithm to normalize string data such as the company name and location entered in text input. This terminal extracts numerical values ​​from specific cell ranges from the uploaded spreadsheet file and expands them in memory as an array of floating-point numbers. This terminal converts the historical certificates entered as image files into multidimensional tensors of pixel values. This terminal serializes these converted data into a hierarchical data format consisting of key-value pairs. This terminal sends the serialized data to the server via an encrypted communication protocol. This terminal performs validation processing on the client side to detect missing or outlier values ​​in the input data beforehand. For example, if the input value for capital is negative, or if the establishment date is a future date, this terminal immediately renders an error message. This terminal prevents invalid data from being sent to the server through prior verification on the client side. This terminal reduces network bandwidth consumption and lowers the overall system communication load. This terminal parses the input website URL and verifies its validity by querying the Domain Name System (DNS). If the URL is determined to be valid, the terminal starts a background process to asynchronously retrieve the website's initial metadata.This terminal sends the acquired metadata to the server as part of the input data. Through these processes, the terminal not only reduces the burden of manual input but also brings about technological improvements in the accuracy and efficiency of data collection as a computer system. A specific application area for this system is the supplier onboarding process in the purchasing departments of manufacturing companies that need to continuously evaluate a large number of new suppliers.

[0041] The input support unit on the terminal side has functions such as automatically suggesting previously registered company names or similar company names when a company name is entered, prompting a list of past representatives when a representative name is entered, or automatically completing the information based on the postal code when a location is entered. The file upload unit automatically determines the format of the uploaded file and extracts text from PDFs and performs OCR processing from images as needed to achieve data standardization. The user authentication unit has security functions such as single sign-on and multi-factor authentication, and controls access permissions by department. Specifically, whenever a user enters text in an input field, this input support unit sends a query to the search application programming interface on the server side using asynchronous communication. This input support unit receives search results based on an N-gram string similarity calculation algorithm or edit distance calculation algorithm executed on the server side. This input support unit dynamically renders the received list of similar company names on the screen in a pre-drop-down format. This file upload unit reads the magic number contained in the header information of the file and accurately determines the file format regardless of the extension. When an image file is entered, this file upload unit activates an optical character recognition model using a convolutional neural network. This optical character recognition model receives a two-dimensional image tensor with resolution normalized and grayscale converted as input data. As output data, it outputs the coordinates of the bounding boxes of the recognized characters and the confidence score for each character. For character regions with confidence scores below a preset threshold, the file upload department performs subsequent processing, assigning a flag that prompts manual confirmation from the user. The user authentication department uses a secure assertion markup language or open authentication protocol to interact with external identity providers. The user authentication department securely stores the encrypted token generated upon initial login in local storage. In subsequent communications, the user authentication department achieves stateless authentication by attaching this token to the header of the Hypertext Transfer Protocol. The user authentication department maintains access control lists for each department in memory and dynamically switches the display and non-display of operational elements on the screen based on the user's role. Through these advanced input support and automated processing, the system physically eliminates human error during data input. The system significantly improves the processing speed of converting unstructured data into structured data through high-precision text extraction based on the optical character recognition model. This system fundamentally improves the efficiency of computer data processing. A specific application area for this system is the document management system used in legal and compliance departments that handle large volumes of contracts and certificates daily.

[0042] Information input from the terminal is sent to the information acquisition department on the server side. In addition to the information received from the terminal, the information acquisition department automatically collects relevant information from multiple internal company databases, including existing customer databases, contract databases managed by the legal department, financial databases managed by the accounting department, transaction history databases managed by the sales department, and audit record databases managed by the audit department. For example, to confirm with existing customers, a search is performed using keys such as company name, representative name, location, and establishment date to automatically retrieve whether the same or similar customers have registered in the past. Furthermore, through external coordination, relevant information is automatically obtained from databases of credit investigation agencies, publicly available information from administrative agencies, corporate information websites, news websites, social networking sites (SNS), industry association lists, and patent information databases. For example, past financial scores and credit ratings are collected from credit investigation agencies; administrative guidance records and law violation records are collected from administrative agencies; corporate reputation and word-of-mouth, news reports, and online criticism history on SNS are collected from the internet; membership status is collected from industry associations; and intellectual property ownership status is collected from patent information databases. Specifically, this information acquisition department parses the received input data and dynamically generates search queries required for querying various databases. When accessing internal company databases, this department utilizes a distributed search engine to perform parallel searches across multiple data sources. To absorb differences in company name representation, this department uses a natural language processing model to convert company names into normalized vector representations. By calculating cosine similarity on a vector database, this department not only extracts completely identical existing customers with high precision but also extracts semantically similar existing customers with high precision. This department queues asynchronous data acquisition tasks for external collaboration departments. This external collaboration department retrieves tasks from a message queue and issues Hypertext Transfer Protocol (HTTP) requests to the application programming interfaces (APIs) of external information sources. When retrieving information from websites, this external collaboration department controls a headless browser to render dynamically generated document object models and extracts text data based on specified XPath or Cascading Style Sheets (CSS) selectors. This external collaboration department applies named entity extraction algorithms to the acquired unstructured text data, extracting entities such as company names, person names, dates, and amounts as structured data. This External Liaison Department receives real-time streaming data containing submissions with specific keywords while acquiring data from social networking services. The department applies a sentiment analysis model to the received streaming data to calculate the probability distributions of positive, negative, and neutral sentiment. The Information Acquisition Department merges massive amounts of data collected from both internal and external sources into a comprehensive database, using unique enterprise hash values ​​as keys. Through these parallel distributed processing and vector search techniques, the system performs advanced name-and-address merging processes at high speed, which was impossible with simple string searches in traditional relational databases. This system delivers a dramatic improvement in the comprehensiveness and speed of data collection.Specific application areas of this system include due diligence of investment target companies in investment funds and comprehensive management systems for global customer networks in general trading companies.

[0043] The vast amount of information collected by the Information Acquisition Department is sent to the Judgment Department on the server side. The Judgment Department uses generative AI to analyze and judge various types of information. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profits, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, increases and decreases in fixed assets, and the content of investment activities, etc., and compares them with industry averages or past customer data to score the soundness. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extracting the legitimacy of legal entity status, the history of changes in representatives, the content recorded in the purpose column, the history of increases and decreases in capital, the history of changes in location, records of dissolution or liquidation, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection systems, the generative AI analyzes the input checklists or regulatory documents, automatically determining their suitability with laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, the existence of countermeasures against information leakage, and the status of obtaining third-party certifications, etc. Regarding past business performance and administrative guidance records, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, content of administrative guidance, existence of improvement reports, implementation status of countermeasures to prevent recurrence, etc., and assesses the risk level. Regarding the content recorded on the company's website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, differences from past records, accuracy of product and service descriptions, whether there are any violations of laws and regulations, and whether there is any infringement of third-party rights. Regarding alliance partner information, generative AI analyzes the credit information, past transaction performance, reputation in the industry, appropriateness of cooperation content, contract content, etc. of partner companies to assess risks and synergies. Regarding business plans, generative AI automatically summarizes the content of business plans, analyzes the appropriateness of sales targets, the realism of investment plans, the feasibility of growth strategies, the comprehensiveness of risk analysis, comparison with similar past cases, and integration with industry trends, etc. For example, it warns of risks when sales targets deviate significantly from past performance or industry averages, and prompts a reassessment of the appropriateness of fundraising plans when investment plans are too large. Specifically, this decision-making unit executes a large-scale language model based on the Transformer architecture and a multimodal neural network. As input data, this decision-making unit receives a 128-dimensional numerical vector normalized from financial statements. As input data, this decision-making unit receives a 512-dimensional embedding vector, extracted from all historical documents through byte-pair encoding. As input data, this decision-making unit receives a 256-dimensional feature tensor, processed from screenshots of the company's website using a convolutional neural network. This decision-making unit combines these heterogeneous data and uses a self-attention mechanism to calculate the hidden correlations between the data.This judgment unit outputs a continuous score representing financial soundness from 0.0 to 1.0. It also outputs a percentage value indicating the probability of a compliance violation. Furthermore, it outputs a one-hot encoded vector representing the classification of risk factors. If the output score exceeds a pre-set threshold, the judgment unit performs branch processing to classify the enterprise as high-risk. This judgment unit then forwards the data of enterprises classified as high-risk to another module for further processing requiring detailed manual review. The judgment unit receives text data from business plans as input and uses an encoder-decoder model to output a summary text string extracting important context. This system extracts nonlinear relationships in high-dimensional space using neural networks. This system identifies potential risks that were previously difficult to detect through manual, rule-based reviews with high speed and accuracy. This system is not simply about automating business reviews, but rather about improving the accuracy of data parsing through computers. Specific application areas of this system include credit assessment of financing targets in banks and business assessment of startups that are investment targets in venture capital.

[0044] The judgment department scores each item individually and performs a comprehensive suitability assessment. For example, if financial soundness falls below a certain standard, there is a significant past record of administrative intervention, false information is found on the website, the personal information protection system is inadequate, risks are identified with alliance partners, or the business plan is unrealistic, the item is deemed unsuitable, and the reasons are automatically generated and notified to the client. Conversely, if all items meet the standards, the item is deemed suitable for customer registration, and registration is processed. The judgment result, along with detailed scoring results, supporting data, AI summary comments, and comparisons with similar past cases, is sent to the results display unit on the client side. Specifically, to integrate multiple individual scores output from various analysis modules, this judgment department executes an ensemble learning model using a gradient boosting decision tree algorithm. The input data for this judgment department includes a 10-dimensional numerical vector representing financial scores, compliance scores, website reliability scores, and personal information protection scores. The input data for this judgment department also includes an array of weighted parameters representing the importance of each item. This judgment department processes this input data according to the conditional branches in the nodes of the decision tree. This decision-making unit outputs a binary label indicating whether the final transaction is feasible. It also outputs an integer score from 0 to 100 representing the overall risk level. Furthermore, it outputs an importance vector representing the contribution of each feature in the decision tree. If the binary label indicates an unsuitable value, the unit performs subsequent processing on the feature with the highest contribution in the importance vector. Based on the output features, the unit dynamically generates a large-scale language model's prompt. This prompt is input into the large-scale language model, causing it to output a natural language text string explaining the specific reasons for unsuitability. If the decision is deemed suitable, the unit issues a transaction commit command to the database management department, inserting a new record into the client's master table. The unit serializes the decision result dataset into JSON format and sends it to the terminal via Hypertext Transfer Protocol. This system utilizes an ensemble approach combining multiple machine learning models to eliminate overfitting or bias that might occur in a single model. The system stably outputs objective and reproducible decision results. This system achieves rigorous data-driven decision-making independent of human subjectivity, significantly improving the overall reliability of the system. As specific application areas of this system, examples include tenant occupancy verification systems in real estate management companies and verification systems for new store owners in e-commerce platforms.

[0045] Registered customer information is automatically categorized by the server-side mapping department based on multiple attributes, including region, industry, transaction size, risk level, alliance partner attributes, business plan content, administrative guidance record, intellectual property ownership status, number of employees, capital size, and years of establishment. For example, it automatically generates enterprise maps based on various entry points, such as IT companies in the Kanto region with annual revenue exceeding 1 billion yen and low risk levels, companies with past administrative guidance records but now deemed to have improved, growing companies with substantial intellectual property holdings, and startups less than 3 years old. The mapping department displays this information using various visualization methods, including graph structures, maps, heatmaps, network diagrams, tree diagrams, and timelines. The results are displayed on the terminal side, allowing department heads to refer to the information in real time. For instance, when discussing new cases, sales department heads can quickly grasp the distribution of existing customers in the same region and industry, the positioning of high-risk companies, the network structure of alliance partners, and the time-series progression of administrative guidance records. Specifically, this mapping unit executes a manifold learning-based dimensionality reduction algorithm to process multidimensional attribute data extracted from the database. The input data for this mapping unit includes a 50-dimensional feature vector group representing the numerical values ​​of each enterprise's attributes. The input data also includes an adjacency matrix representing transaction or capital relationships between enterprises. Using the dimensionality reduction algorithm, this mapping unit outputs two-dimensional or three-dimensional coordinate data while preserving the local structure between data points in the high-dimensional space. The output data includes floating-point numbers representing the X and Y coordinates of each enterprise node. The output data includes a numerical array representing the edge weights between nodes. The output data includes integer values ​​for group identifiers assigned by the clustering algorithm. Using the calculated coordinate data and group identifiers, this mapping unit performs subsequent processing to generate a graph data structure interpretable by the front-end's drawing library. This mapping unit sends the generated graph data to the terminal. The terminal's result display unit uses a network graphics library to render the received graph data at high speed with hardware acceleration. The result display unit dynamically adjusts the resolution and level of detail of the displayed nodes based on the user's zoom or pan operations. This system provides visually understandable information by mapping high-dimensional and complex data to a low-dimensional space using advanced computational processing. Compared to previous interfaces that relied on reading large amounts of text data, this system offers a dramatic reduction in cognitive load and a significant improvement in decision-making speed. Specific applications of this system include customer segmentation visualization in marketing departments and analysis systems for alliance networks with competitors in business planning departments.

[0046] Furthermore, the server-side AI learning management department continuously learns from past judgments made by the judgment department, newly acquired external information, the occurrence of disputes after actual transactions, the occurrence of administrative guidance, customer bankruptcies or mergers, growth performance, etc., to improve the judgment accuracy of generative AI. For example, if a company previously judged as unsuitable later causes problems, this case is imported as learning data and reflected in future judgment criteria. Conversely, cases previously judged as high-risk but which did not actually cause problems are also utilized in the re-evaluation of judgment criteria. In the event of new legal amendments or changes in industry guidelines, the AI ​​learning management department also automatically acquires information and reflects it in the judgment department's algorithm. As a result, the overall judgment accuracy and reliability of the system are continuously improved. Specifically, this AI learning management department constructs a dedicated learning pipeline independent of the inference model running in the production environment. As input data, this AI learning management department takes in a teacher dataset consisting of pairs of input feature vectors from past judgments and correct labels based on subsequent actual events. As input data, this AI learning management department takes in an embedding tensor after vectorizing newly acquired legal texts. This AI learning management unit divides the data into mini-batches and performs forward propagation processing on the neural network. As output data, this unit outputs a scalar value of the loss function representing the error between the model's predicted value and the correct label. As output data, this unit also outputs gradient vectors for the weight parameters of each layer. Using the calculated gradient vectors, this unit performs subsequent processing to update the weight parameters using stochastic gradient descent or adaptive moment estimation algorithms. The updated model is evaluated using a validation dataset, and only if the accuracy metric exceeds that of the previous model is a seamless blue-green deployment replacing the production model. This unit uses a reinforcement learning framework, designing a reward function that provides positive rewards for decisions that bring commercial benefits and negative rewards for decisions that bring losses, optimizing the model through policy gradient methods. This system incorporates an online learning mechanism that autonomously adapts to changes in the environment or new data tendencies. This system effectively prevents model staleness and maintains high inference accuracy over the long term. This system achieves a self-evolving computer system, impossible for static rule-based systems. Specific applications of this system include the continuous updating of violation detection systems in credit card companies and the use of dynamic premium calculation models in insurance companies.

[0047] The server-side database management department centrally manages all customer information, judgment results, scoring data, mapping data, AI learning data, user operation logs, external interaction history, etc., and saves them as audit trails. The log management department meticulously records all operation history, data acquisition history, judgment history, mapping history, etc., for future reference by the audit or compliance departments. The security management department installs data encryption, access control, unauthorized access detection, data backup, and recovery functions in case of failure to ensure the overall security and reliability of the system. Specifically, this database management department adopts a multi-language persistence architecture that combines a relational database for storing structured data with a distributed non-relational database for storing unstructured or time-series data. This database management department caches frequently accessed data in an in-memory data grid, reducing disk I / O overhead. This log management department receives log messages asynchronously output from various modules through a message broker and writes them to a unified log storage via a log collection broker. This log management department applies hash chaining technology to log data, mathematically detecting data tampering by including the hash value of each log entry in the next entry. This security management department encrypts all stored data using a 256-bit key length and an Advanced Encryption Standard (AES) algorithm. It applies Transport Layer Security (TLS) protocols to data along communication paths. This department continuously monitors network traffic, using machine learning-based anomaly detection algorithms to identify data packet flows that deviate from typical communication patterns in real time. Upon detecting anomalies, this department automatically blocks subsequent communication from the corresponding Internet Protocol (IP) address via firewall. This department periodically takes snapshots of the database and performs asynchronous replication to a geographically distant data center. Through these advanced data management and security mechanisms, this system eliminates single points of failure, maximizing system availability and data integrity. This system provides technological improvements to enhance fault tolerance against network attacks or hardware failures. Specific application areas for this system include secure electronic medical record management systems in healthcare institutions and comprehensive database systems for confidential information in government agencies.

[0048] On the terminal side, the user interface provides comprehensive input support, auto-completion, file upload, result display, notification, download, and user authentication functions. For example, it clearly displays suitable judgment results, reasons for unsuitability, visual information of the enterprise map, AI summary comments, risk alerts, and requests for additional information sent from the judgment department, and allows downloading in PDF or Excel format as needed. The notification department notifies the user in real time of judgment results, requests for additional information, risk alerts, and feedback on AI learning results. Specifically, this user interface is built using a component-oriented front-end framework. Running as a single-page application, it does not reload the entire screen but uses a virtual document object model to efficiently re-render changed parts. This user interface maintains JSON-formatted data received from the server in local state management storage and synchronizes data across components. Based on the judgment result score, the user interface dynamically changes the color or icon of the indicator on the screen. The notification department establishes a WebSocket connection with the server to maintain a bidirectional real-time communication channel. When this notification unit receives an event message pushed from the server, it invokes the operating system's native notification application programming interface (API) to display a pop-up notification on the user's desktop, bypassing the browser's frame. This download unit receives data from the server as a binary large object and encodes it into a specified file format on the client-side storage. The download unit generates a virtual Uniform Resource Locator (URI) for the encoded data and initiates the browser's download manager to save it to local storage. Through these asynchronous processes and client-side rendering optimizations, this system minimizes the impact of network latency. This system significantly improves the system's responsiveness to user operations, providing a smooth user experience. Specific application areas for this system include real-time trading dashboards in securities companies and dynamic delivery status monitoring systems in logistics networks.

[0049] In this embodiment, the main functions such as the information acquisition department, judgment department, mapping department, AI learning management department, database management department, external linkage department, log management department, and security management department are implemented on the server side, while the user interface department, input support department, result display department, notification department, file upload department, download department, and user authentication department are implemented on the terminal side. Generative AI, as an essential component, is assembled in the judgment department and AI learning management department on the server side and is applied to various information processing tasks such as parsing, judgment, learning, summarizing, comparing, and visualizing. For example, by combining the diverse functions of generative AI, such as document parsing based on natural language processing, extraction of certificate content based on image recognition, risk assessment based on anomaly detection, automatic summarization of business plans based on summary generation, prompts for judgment criteria based on similar case searches, automatic capture and summarization of external information, trend analysis of time series data, and visualization of alliance structures based on network analysis, high-precision and efficient customer information management that was previously difficult to achieve manually can be realized. Specifically, this system adopts a client-server architecture that clearly separates the lightweight front-end processing on the terminal side from the heavyweight back-end processing on the server side. This system loosely connects processing modules via message queues, constructing a data processing pipeline. As input data flows sequentially from the information acquisition unit to the decision unit and mapping unit, the system invokes generative AI inference endpoints at each stage. The system executes generative AI inference processing in parallel on a cluster of image processing devices, maximizing throughput. The system uses natural language processing models to extract semantic feature vectors from unstructured text and image recognition models to extract spatial feature tensors from pixel data. It integrates these heterogeneous feature quantities in a combination layer and uses a multilayer perceptron to calculate the final decision score. In trend analysis of time-series data, the system uses long short-term memory networks to predict future values ​​from past data sequences. In network analysis, the system uses graph convolutional networks to quantify the strength of relationships between nodes. By integrating these advanced computational processes, the system assesses complex business risks from multiple perspectives that cannot be fully captured by a single algorithm. This system processes data at a scale and speed exceeding the limits of human cognitive abilities, bringing fundamental improvements to business processes due to the advancement of computer technology itself. As specific areas of application for this system, examples include global compliance integration management platforms in multinational corporations and suitability review systems for large-scale public procurement in government agencies.

[0050] (System Composition) The system described in this implementation includes an information acquisition department, a judgment department, a mapping department, a database management department, an external linkage department, an AI learning management department, a log management department, a security management department, a user interface department, an input support department, a result display department, a notification department, a file upload department, a download department, and a user authentication department. The information acquisition department receives data in various formats from terminals, including the company name, representative name, location, date of establishment, capital, number of employees, business content, financial statements for the past three periods, all historical certificates, information regarding the personal information protection system, past business performance, administrative guidance records, company website URL, alliance information, business plans, and geographic information of potential new customers. It also automatically collects relevant information from multiple internal company databases, such as the existing customer database, the legal department's contract database, the accounting department's financial database, the sales department's transaction history database, and the audit department's audit record database. For example, using the company name, representative name, location, and date of establishment as keys, it automatically extracts duplicates with existing customers or references past contract content and transaction history. Furthermore, in collaboration with external departments, the department obtains past financial scores and credit ratings from credit investigation agency databases, administrative guidance records and law enforcement violation records from publicly available information from administrative agencies, and automatically collects reputation, word-of-mouth, news reports, online criticism history, membership status, and intellectual property holding status from online corporate websites, news websites, social networking sites, industry group lists, and patent information databases. For example, it also analyzes the frequency of company names on social networking sites and their co-occurrence rate with negative terms to assess the risk of online criticism. Specifically, to process the diverse data received, the Information Acquisition Department uses data pipeline orchestration tools to automate a series of extraction, transformation, and loading processes. When collecting information from internal databases, the department uses a data virtualization layer that absorbs the schema differences between various databases to issue standardized structured query languages. For unstructured text data obtained from external departments, the department performs named entity extraction using a pre-learned language model. The department constructs a knowledge graph in memory, with extracted entities as nodes and their relationships as edges. This information acquisition department applies a sliding window streaming algorithm to data streams from social networking services. It counts the co-occurrence frequency of company names with words carrying negative sentiment within a specific timeframe, using statistical significance tests to detect sudden online criticism events. The department also analyzes patent classification codes and citation data from patent information databases to calculate indicators that quantify a company's technological competitiveness. This system integrates information from these diverse data sources in real-time and asynchronously. This system effectively eliminates data silos and builds a single, reliable information source.This system breaks through the limitations of manual information collection, achieving comprehensive and high-speed data aggregation based on computers. Specific applications of this system include comprehensive analysis systems for target companies in M&A consulting and underwriting review systems for enterprise-oriented cyber insurance in insurance companies.

[0051] The judgment department inputs the vast amount of information collected by the information acquisition department into the generative AI to perform various information analysis and judgments. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profits, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, the increase or decrease of fixed assets, the content of investment activities, etc., and compares them with industry averages or past customer data to score the soundness. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extract the legitimacy of legal person status, the history of changes in representatives, the content recorded in the purpose column, the history of increases or decreases in capital, the history of changes in location, dissolution or liquidation records, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection systems, the generative AI analyzes the input checklists or regulatory documents, automatically determines their suitability with laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, whether there are countermeasures against information leakage, and the status of obtaining third-party certifications, etc. Regarding past business performance and administrative guidance records, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, content of administrative guidance, existence of improvement reports, implementation status of countermeasures to prevent recurrence, etc., and assesses the risk level. Regarding the content recorded on the company's website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, differences from past records, accuracy of product and service descriptions, whether there are violations of laws, and whether there is infringement of third-party rights, etc. Regarding alliance partner information, generative AI analyzes the credit information, past transaction performance, reputation in the industry, appropriateness of cooperation content, contract content, etc. of partner companies to assess risks and synergies. Regarding business plans, generative AI automatically summarizes the content of business plans, analyzes the appropriateness of sales targets, the realism of investment plans, the feasibility of growth strategies, the comprehensiveness of risk analysis, comparison with similar past cases, and consistency with industry trends, etc. For example, it warns of risks when sales targets deviate significantly from past performance or industry averages, and prompts a reassessment of the appropriateness of fundraising plans when investment plans are too large. The assessment department scores each of these items separately, performs a comprehensive suitability assessment, generates detailed scoring results, supporting data, AI-based summary comments, comparisons with similar past cases, etc., and sends them to the results display department on the terminal side.Specific examples of prompts for generative AI to read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal status and past risk factors from the contents of all certificates for the following historical matters, and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following company websites and determine whether there are any false records or exaggerated advertisements," and "Summarize the following business plan, compare it with similar past cases, and analyze its growth potential and risks." Specifically, this judgment department uses prompt engineering technology to dynamically combine task-optimized system prompts and user prompts. As input data, this judgment department takes structured data in JSON format, with each item in the financial statements as the key and numerical values ​​as the value. As input data, this judgment department takes the text of the business plan as input, dividing it into chunks and converting each chunk into a 768-dimensional vector embedding tensor. As input data, this judgment department takes the top 5 text data extracted from a database of similar past cases through vector similarity search as context input. This decision-making unit inputs the input data into the encoder of a large-scale language model, calculating the dependencies between the tags through a self-attention mechanism. As output data, this decision-making unit outputs an array of integer values ​​from 0 to 10 representing the scores of each evaluation item. As output data, this decision-making unit also outputs binary log-probability values ​​representing the presence or absence of risk. Finally, this decision-making unit outputs a Markdown-formatted text string explaining the basis for the decision. This decision-making unit applies a Sigmoid function to the output log-probability values, performing subsequent processing to convert them into probability values. If the probability value exceeds a set threshold, this decision-making unit sets a risk alarm flag to true and triggers an event to the notification unit. This system, by leveraging the reasoning capabilities of a large-scale language model, accurately interprets ambiguous text expressions and complex contexts that were previously impossible to handle with rule-based systems. This system achieves refined scoring based on advanced natural language understanding, resulting in a significant improvement in the accuracy of decision-making processes. Specific applications of this system include automated support systems for internal control assessments in auditing firms and due diligence systems for business operations in consulting companies.

[0052] The Mapping Department automatically categorizes companies based on various attributes, including the judgment and risk assessment results from the Judgment Department, the company's region, industry, transaction size, risk level, alliance partner attributes, business plan content, administrative guidance records, intellectual property holdings, employee size, capital size, and years of establishment. It then generates company maps using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, and timelines. These maps are displayed on the terminal, allowing department heads to refer to them in real time. For example, the department can automatically generate company maps from various entry points, such as IT companies in the Kanto region with annual revenue exceeding 1 billion yen and low risk levels, companies with past administrative guidance records but now deemed to have improved, growing companies with substantial intellectual property holdings, and startups less than three years old. When business department heads discuss new projects, the Mapping Department provides a clear overview of the distribution of existing clients in the same region and industry, the positioning of high-risk companies, the network structure of alliance partners, and the timeline of administrative guidance records. Specifically, this mapping unit executes a graph neural network to model the complex relationships between enterprises. The input data for this mapping unit is a graph structure with enterprises as nodes and transaction and capital relationships between enterprises as edges. The input data for this mapping unit also includes 64-dimensional feature vectors representing the attribute information of each enterprise. This mapping unit uses graph convolutional layers to aggregate the feature values ​​of neighboring nodes and update the hidden state vectors of each node. The output data for this mapping unit is a new embedding vector for each node reflecting the structural features of the entire graph. The output data for this mapping unit is a score matrix representing the probability of the existence of potential links between nodes, ranging from 0.0 to 1.0. The output data for this mapping unit is an integer value for the cluster ID assigned by the community detection algorithm. This mapping unit applies a unified manifold approximation and projection algorithm to the computed embedding vectors to perform subsequent processing to generate two-dimensional coordinate data for depiction. Using the generated coordinate data and cluster IDs, this mapping unit constructs a JSON-formatted visualization dataset that can be drawn on the front end. This system mathematically extracts potential hidden risk propagation paths and synergistic effects between enterprises that cannot be discovered by attribute filtering alone, through advanced relational learning based on graph neural networks. By transforming high-dimensional network structures into intuitive visual representations, this system expands human information processing capabilities, leading to technological improvements that support rapid and accurate decision-making. Specific applications of this system include systems for bottleneck enterprises in supply chain management and simulation systems for cascading failure risks in financial institutions.

[0053] The Database Management Department centrally manages all customer information, judgment results, scoring data, mapping data, AI learning data, user operation logs, and external linkage history, and saves them as audit trails. For example, it saves all previously registered customer information, scoring results output by the Judgment Department, enterprise maps generated by the Mapping Department, learning data accumulated by the AI ​​Learning Management Department, all user operation history, and the acquisition history of external data obtained by the External Linkage Department in a time-series format, allowing for searching, retrieval, and reuse as needed. Specifically, this Database Management Department has a routing mechanism that dynamically selects the optimal storage engine based on the characteristics of the data. For customer master data requiring transactional consistency, this Database Management Department uses a relational database engine that guarantees ACID properties. For massive log data and time-series scoring data, this Database Management Department uses a columnar time-series database engine dedicated to write performance. This Database Management Department serves as input data for data insertion and update queries asynchronously sent from various modules. This Database Management Department also serves as input data for search queries containing complex join conditions sent from the User Interface Department. This database management unit executes the query optimizer to generate the optimal execution plan, scanning the index tree for specific target records. As output data, this database management unit outputs a result set representing the collection of records matching the search criteria. As output data, this database management unit outputs performance metrics such as query execution time and the number of pages read. As output data, this database management unit outputs status codes indicating the success or failure of transactions. This database management unit stores the result sets of frequently executed search queries in a memory cache server, reducing the response time for the next execution of the same query to milliseconds. Through a combination of distributed database architecture and memory caching, this system maintains high-speed read and write performance even for petabyte-scale datasets. This system effectively balances data persistence and search efficiency. Specific applications of this system include integrated management systems for customer and product data in large-scale e-commerce platforms and real-time analysis systems for massive call history data in telecommunications operators.

[0054] The External Liaison Department collaborates with multiple external information sources, including credit investigation agencies, government agencies, corporate websites, news websites, social networking sites (SNS), industry associations, and patent databases, using methods such as APIs, web scraping, RSS subscriptions, and database integration to automatically acquire necessary information. For example, it obtains credit ratings and financial scores from credit investigation agencies via API integration; administrative guidance records and law violation records from government agencies via web scraping; analyzes the frequency of company names and their co-occurrence with negative terms from SNS; obtains membership qualifications and industry reputation from industry associations; and acquires the ownership and application status of intellectual property rights from patent databases. Specifically, this External Liaison Department uses an asynchronous input / output framework to efficiently manage simultaneous connections to thousands of external endpoints. It dynamically switches between RESTful Hypertext Transfer Protocol (HTTP) requests and gRPC binary communication based on the specifications of the target external information sources. During web scraping, it automatically parses the target website's bot exclusion protocol and executes rate-limiting algorithms to control access frequency. This external linkage unit takes as input a dictionary object containing search parameters including the company name and identifier of the target. It also takes as input strings the authentication tokens and API keys of various external information sources. Using this input data, the external linkage unit constructs the request header and body and sends them via network sockets. As output data, it outputs the response payload in JSON or XML format returned from the external server. It also outputs the Hypertext Transfer Protocol (HTTP) status code and a numerical value representing the communication latency in milliseconds. In case of an error in the status code, the external linkage unit performs subsequent processing, including retrying according to the exponential backoff algorithm. The external linkage unit parses the acquired response payload, maps it to the unified data model within the system, and forwards it to the information acquisition unit. Through these advanced communication controls and error handling mechanisms, this system achieves stable data collection unaffected by the instability of the external network or fluctuations in the load of the target server. This system brings technological improvements to enhance fault tolerance in interactions with external systems. As specific applications of this system, examples include real-time market data aggregation systems in financial information service providers and seat availability information integration systems from multiple airlines in travel agencies.

[0055] The AI ​​Learning Management Department continuously learns from the judgment results made by the Judgment Department in the past, new information acquired from external sources, whether disputes occurred after actual transactions, whether administrative guidance was issued, whether clients went bankrupt or merged, and their growth performance, to improve the accuracy of generative AI judgments. For example, if a company previously judged as unsuitable later causes problems, this case is imported as learning data and reflected in future judgment benchmarks. In the event of new legal amendments or changes in industry guidelines, the AI ​​Learning Management Department also automatically acquires information and reflects it in the judgment department's algorithm. Specifically, this AI Learning Management Department implements an MLOps infrastructure that integrates the lifecycle management of machine learning within the continuous integration and continuous delivery pipeline. The input data for this AI Learning Management Department consists of a tensor dataset composed of tens of thousands of records that pair input features from past inference with correct answer labels based on subsequent observations. The input data for this AI Learning Management Department includes an array of initial values ​​for hyperparameters such as the learning rate and batch size of the model. The input data for this AI Learning Management Department includes a context vector encoded using a natural language processing model from newly acquired legal text. This AI learning management department uses this data to perform weight update processing using a neural network-based error backpropagation method. As output data, this AI learning management department outputs a scalar value of the cross-entropy error for the validation dataset. As output data, this AI learning management department outputs an array of floating-point numbers representing precision, recall, and F1 score. As output data, this AI learning management department outputs a binary file containing the updated model's weight parameters. This AI learning management department performs subsequent processing to verify whether the calculated F1 score statistically significantly exceeds the score of the current production model. If a significant improvement is confirmed, this AI learning management department pushes the new model image to the container registry and issues a rolling update command to the orchestration tool. This system automates the entire process from model relearning to evaluation and deployment without human intervention. This system autonomously adapts to changes in data distribution (data drift) and changes in concepts (concept drift), maintaining optimal inference accuracy. Specific applications of this system include automatic update systems for personalized product recommendation models in e-commerce websites and continuous improvement systems for defect detection models in manufacturing production lines.

[0056] The Log Management Department meticulously records all operation history, data acquisition history, judgment history, and mapping history, enabling the audit and compliance departments to refer to them later. For example, it records in chronological order who input which customer information, when, what judgment result was output, when which external information was acquired, and what mapping was generated, and saves this as an audit trail. Specifically, this Log Management Department executes a log forwarding agent, which collects unstructured log messages asynchronously output from various microservices within the system and converts them into structured data. The Log Management Department receives JSON-formatted log entries containing timestamps, event types, user identifiers, target resource identifiers, and payloads as input data. It also receives Syslog-formatted text strings output from network devices as input data. The Log Management Department uses a regular expression engine to extract necessary fields from the text strings and maps them to key-value pairs. As output data, the Log Management Department outputs indexed document data optimized for search engines. Finally, it outputs time-series metrics representing the log data volume and error rate. This log management department outputs anomaly scores for each log entry, calculated by anomaly detection algorithms. When anomaly scores exceed a set threshold, the department sends an immediate alert to the security management department. The department applies unsupervised machine learning algorithms, such as the Isolation Forest algorithm, to the accumulated log data, automatically identifying suspicious accesses that deviate from normal operating patterns. This system not only stores massive amounts of log data but also performs real-time searches and advanced analysis. It provides technological improvements to support early detection of security incidents and rapid root cause analysis of system failures. Specific applications of this system include access log analysis systems for insider trading monitoring in financial institutions and large-scale infrastructure monitoring systems in cloud service providers.

[0057] The Security Management Department implements data encryption, access control management, unauthorized access detection, data backup, and fault recovery functions to ensure the security and reliability of the entire system. For example, it meticulously sets access permissions for users, controls viewable and editable information by department, encrypts all communications, automatically issues alerts upon detecting unauthorized access, regularly performs data backups, and enables rapid recovery in case of failure. Specifically, this Security Management Department implements an access control mechanism based on a zero-trust architecture that does not rely on network-dependent perimeter defense. The Security Management Department inputs authentication information (password hash and one-time password) sent during a user's login attempt. It also inputs a context vector containing the Internet Protocol address of the access source, the device fingerprint, and location information. Furthermore, it inputs the identifier of the requested resource and the operation type (read, write, delete, etc.). This input data is fed into a risk-based authentication engine to calculate the deviation from past access patterns. The Security Management Department outputs a Boolean value indicating successful or failed authentication. As output data, this security management department outputs authorization decisions and policies indicating whether to grant or deny access requests. As output data, this security management department also outputs a JSON Web token containing the session's validity period and permission scope. In the event of authorization denial, this security management department performs follow-up actions by logging the access attempt in the security information event management system and displaying a warning on the administrator's dashboard. During database backup processing, this security management department extracts only the differential data, compresses it, encrypts it, and asynchronously forwards it to remote object storage. This system provides robust protection against internal threats and advanced external cyberattacks through a combination of multi-layered security defenses and dynamic risk assessment. This system effectively minimizes the risk of information leakage and ensures the system's business continuity. Specific applications of this system include protection systems for classified technical information in the defense industry and personal health information management systems for patients in the healthcare industry.

[0058] The user interface department operates on the terminal side, comprehensively providing input support, auto-completion, file upload, result display, notification, download, and user authentication functions. For example, when a company name is entered, it automatically displays a list of past candidate names or similar company names; when uploading financial statements, it automatically performs format conversion; and it displays easily understandable information such as suitable and unsuitable judgment results sent from the judgment department, visualizations of the company map, AI-based summary comments, risk alerts, and requests for additional information, and allows downloading in PDF or Excel format as needed. The notification department notifies users of judgment results, requests for additional information, risk alerts, and feedback on AI learning results in real time. The file upload department automatically determines the format of uploaded files, extracting text from PDFs and performing OCR processing from images as needed to standardize the data. The download department enables the download of judgment results, company maps, and AI-based summary comments in various file formats. The user authentication department has security functions such as single sign-on and multi-factor authentication, and controls access permissions by department. Specifically, this user interface department adopts an architecture that offloads heavy computational processing to background threads called Web Workers to avoid blocking the browser's main thread. This user interface unit takes as input data event objects such as user keyboard input and mouse clicks. It also takes as input data the JSON payload of the judgment results asynchronously received from the server. Furthermore, it takes as input array data of the nodes and edges of the graph structure received from the mapping unit. This user interface unit uses a state management library to integrate this input data into a unified state tree, detecting data changes and recalculating only the relevant screen components. As output data, this user interface unit outputs the node tree of the virtual document object model (VDOM) to the browser's rendering engine. It also outputs an array of pixel data drawn on the canvas elements. Finally, it outputs the payload of the Hypertext Transfer Protocol (HTTP) request sent to the server. Based on the differential calculation results of the VDOM, this user interface unit performs subsequent processing, applying minimal patches to the actual document object model. Through these advanced front-end technologies, this system maintains a smooth rendering performance of 60 frames per second, even when handling large amounts of data or complex graphics. This system eliminates user operation latency, resulting in a technological improvement that provides an intuitive and efficient human-computer interface. Specific applications of this system include real-time radar information display consoles used by air traffic controllers and monitoring control panels for manufacturing production lines used by factory operators.

[0059] In this way, the system involved in this implementation achieves high-precision and efficient customer information management centered on generative AI through the coordinated operation of its various components. Specific examples of prompts for the generative AI to read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of the legal entity and past risk factors from the contents of all the following historical certificates and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following corporate website and determine whether there are any false statements or exaggerated advertisements," and "Summarize the following business plan and compare it with similar past cases to analyze growth potential and risks." By using these prompts, the generative AI performs various tasks such as parsing, judging, summarizing, comparing, and visualizing information, achieving high-precision and efficient customer information management that was previously difficult to achieve manually. Specifically, this system is structured around a generative AI module based on a large-scale language model. This generative AI module uses a multi-layer neural network with a Transformer architecture. The system converts the input text data into fixed-dimensional token embedding vectors through a word segmenter. This system performs matrix operations on the vector using a self-attention mechanism. It quantifies contextual dependencies in a high-dimensional space. The system generates an input tensor that combines the prompt text and target data. For example, it formats numerical data from financial statements as a comma-separated time-series array. This time-series array, along with the prompt text, is input as a 512-dimensional input tensor into the generative AI module. Based on pre-learned weight parameters, the generative AI module outputs a continuous score from 0 to 100 representing financial soundness. It also outputs binary labels indicating the presence or absence of risk factors. Finally, it outputs a text vector representing the natural language criteria used for the judgment. The system determines whether the output score exceeds a pre-set threshold. Based on the threshold determination, the system automatically branches subsequent approval processes. This system transforms the subjective document reading and comprehension process, based on human judgment, into a deterministic matrix operation process in a high-dimensional vector space. This system significantly improves the processing speed of review operations. It eliminates biases in review standards, improving judgment accuracy. This system is extremely useful in use cases such as financing reviews in financial institutions and new supplier evaluations in large enterprises. This system periodically performs fine-tuning using the cross-entropy error function. It incorporates the latest legal data and industry trends into the weighting parameters. Over time, the system autonomously improves its judgment accuracy.

[0060] (Implementation steps) Step 1: Input and collection of customer candidate information Users can input or upload files through the user interface of the terminal to provide information such as the company name, representative name, location, date of establishment, capital, number of employees, business scope, financial statements for the past three periods, all historical certificates, information regarding the personal information protection system, past business performance, administrative guidance records, company website URL, alliance information, business plan, and regional information of potential new clients. The input support department performs functions such as automatic completion of company or representative names, postal code conversion, format conversion of financial statements, OCR processing of all historical certificates, checklist creation for the personal information protection system, and template creation of business plans, improving the efficiency of input operations. Specifically, this input support department runs as a dedicated client application deployed on the user terminal's memory. This input support department monitors the user's keystrokes in milliseconds. This input support department sends the input string to the search module on the server side via an asynchronous communication protocol. This search module performs a prefix search using a trie structure on an in-memory database. This search module extracts a list of candidate company or representative names in milliseconds. This search module returns the extracted list as a JSON array to the terminal. This input support unit immediately renders the received array data as a dropdown list on the user interface. This input support unit performs binary header parsing on uploaded file groups. This input support unit deterministically identifies file formats. This input support unit performs optical character recognition processing on image files using convolutional neural networks. This input support unit converts the pixel array within the image into a 256-dimensional feature map. This input support unit extracts the bounding boxes of characters from the feature map. This input support unit converts the extracted regions into text string data. This input support unit converts image files, which are unstructured data, into structured text data that can be interpreted by subsequent natural language processing models. This input support unit extracts the table structure from PDF financial statement files. This input support unit converts the extracted table structure into a multi-dimensional numerical array tensor. This input support unit performs these preprocessing operations in a distributed manner in an edge computing environment. This input support unit significantly reduces the communication and computational load on the server side. This system automates the normalization and structuring of input data. This system physically eliminates human errors during data input. This system improves the inference accuracy in subsequent AI parsing processing. This system eliminates data collection bottlenecks in use cases such as handling a large number of newly opened securities companies and reviewing a large number of affiliated stores of platform operators.

[0061] Step 2: Automatically retrieve information from internal and external databases and external information sources. In addition to information received from terminals, the Information Acquisition Department automatically collects relevant information from the company's existing customer database, legal department's contract database, accounting department's financial database, sales department's transaction history database, and audit department's audit record database. Furthermore, the External Liaison Department automatically acquires credit ratings, administrative guidance records, reputation, word-of-mouth, news reports, online criticism history, membership qualifications, and intellectual property holding status from credit investigation agency databases, publicly available information from administrative agencies, corporate information websites, news websites, social networking sites (SNS), industry group lists, and patent information databases. For example, it also analyzes the frequency of company names on social networking sites or their co-occurrence rate with negative terms to assess the risk of online criticism. Specifically, this Information Acquisition Department controls multiple crawler agents running on a distributed processing framework. Using the target company's identifier as the key, this Information Acquisition Department issues SQL queries or RESTful API requests in parallel to multiple heterogeneous databases. This External Liaison Department collects unstructured text data from the internet. This External Liaison Department performs lexical analysis and named entity extraction on the collected text data. This External Liaison Department extracts text fragments related to the target company. This external linkage unit inputs the extracted text fragments into a word embedding model. This external linkage unit converts the text fragments into 768-dimensional dense vectors. This external linkage unit calculates the cosine similarity between this dense vector and a predefined group of negative word dictionary vectors. This external linkage unit counts the number of text fragments whose cosine similarity exceeds a predetermined threshold. This external linkage unit calculates the proportion of negative text relative to the total number of acquired texts, generating a network reputation risk score as a continuous value ranging from 0.0 to 1.0. This external linkage unit applies an autoregressive moving average model to the frequency of occurrence acquired as time-series data. This external linkage unit predicts future abnormal peaks in frequency of occurrence. This system performs these collection and parsing processes asynchronously through multi-threading. This system minimizes the impact of network latency, ensuring real-time information collection. This system replaces the previous process of manually operating search engines to collect information with an algorithm-based automated data pipeline. This system guarantees the comprehensiveness and objectivity of information collection. This system brings the technical effect of early detection of potential risks in use cases such as reputation risk management in corporate public relations departments and due diligence of target companies in investment funds.

[0062] Step 3: Multi-dimensional information analysis and judgment based on generative AI The judgment department inputs all information collected by the information acquisition department into the generative AI for analysis and judgment of various information. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profits, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, the increase or decrease of fixed assets, the content of investment activities, etc., and compares them with industry averages or past customer data to score the soundness. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extract the legitimacy of legal person status, the history of changes in representatives, the contents recorded in the purpose column, the history of increases or decreases in capital, the history of changes in location, records of dissolution or liquidation, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection systems, the generative AI analyzes checklists or regulatory documents, automatically determines their suitability with laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, whether there are countermeasures against information leakage, and the status of obtaining third-party certifications, etc. Regarding past business performance and administrative guidance records, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, content of administrative guidance, existence of improvement reports, implementation status of countermeasures to prevent recurrence, etc., to assess the risk level. Regarding the content recorded on the company's website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, differences from past records, accuracy of product and service descriptions, whether there are violations of laws, and whether there is infringement of third-party rights, etc. Regarding alliance partner information, generative AI analyzes the credit information or past transaction performance of partner companies, their reputation in the industry, the appropriateness of cooperation content, contract content, etc., to assess risks or synergies. Regarding business plans, generative AI automatically summarizes the content of business plans, analyzes the appropriateness of sales targets, the realism of investment plans, the feasibility of growth strategies, the comprehensiveness of risk analysis, comparison with similar past cases, and consistency with industry trends, etc. Specific examples of prompts for generative AI to read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of the legal entity and past risk factors from the contents of all the following historical certificates, and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following company websites and determine whether there are any false statements or exaggerated advertisements," and "Summarize the following business plans and compare them with similar past cases to analyze growth potential and risks." Specifically, this judgment department possesses an integrated learning architecture that integrates multiple specialized neural network models running on a parallel computing cluster using GPUs. This judgment department categorizes the diverse input data by modality.For structured numerical data such as financial data, this decision-making unit applies a regression model using a multilayer perceptron. It uses financial indicators from the past three periods to form a 128-dimensional input vector. This input vector is fed into the regression model, which outputs a financial soundness score ranging from 0 to 100. For unstructured text data such as prescribed documents or business plans, this decision-making unit applies a large-scale language model based on Transformer. It tokenizes the text data, generating an input tensor with a maximum of 4096 tokens. This input tensor is fed into the large-scale language model. From the hidden state vectors of the final layer of the large-scale language model, it outputs a probability distribution representing the suitability of the legal term. For example, it outputs a suitability probability of 0.85 and an unsuitability probability of 0.15. It then outputs a one-hot vector representing the categories of the extracted risk factors. Combining the scores, probability distributions, and vectors output from each model, it generates a comprehensive feature vector. This comprehensive feature vector is fed into the classifier of the final segment. The decision-making unit outputs a binary decision result representing the overall feasibility of the transaction. This decision-making department sends the output decision results to the database management department, which uses them as triggers for subsequent approval workflows. This system transforms the complex information integration process based on tacit knowledge performed by human reviewers into a mathematical feature synthesis process in a high-dimensional space. This system fully guarantees the objectivity and reproducibility of the review. This system drastically reduces the delivery time required for review from several days to several minutes. This system provides strong technical support for expert judgment in use cases requiring high levels of expertise, such as due diligence of target companies in corporate acquisitions and investment target evaluation in venture capital. This system extracts the attention weights in the reasoning process of each model. This system quantifies which input data contributes most to the decision and visualizes it as the basis for the decision. This system eliminates the black-box problem of AI, achieving interpretable AI.

[0063] Step 4: Determine the results, assign scores, generate and notify supporting information. The judgment department scores each item, performs a comprehensive suitability assessment, and generates detailed scoring results or supporting data, AI-based summary comments, comparisons with similar past cases, etc., which are then sent to the results display department on the terminal side. The notification department notifies the user in real time of the judgment results or reasons for unsuitability, risk alerts, requests for additional information, and feedback on AI learning results. For example, unsuitability is determined in cases such as financial soundness below a certain standard, a significant past administrative guidance record, false information found on a website, inadequate personal information protection mechanisms, risks identified by alliance partners, or unrealistic business plans. Reasons are automatically generated and notified to the terminal side. Specifically, this judgment department inputs the continuous value scores and probability distributions of each item output from the front-end neural network into a decision tree algorithm pre-defined in memory. This judgment department performs threshold assessments at each node of the decision tree. This judgment department deterministically derives the final comprehensive suitability judgment label. For example, this judgment department outputs any of the labels: approval, conditional approval, or rejection. This judgment department records the traversal path of the decision tree. This judgment department identifies the main unsuitability factors based on this path. This decision-making unit inputs identifiers of the key factors into the natural language generation model. It dynamically generates explanatory text corresponding to the identifiers. For example, it generates a string indicating that the equity ratio is 20% lower than the industry average. The decision-making unit packages the score array, decision labels, and generated explanatory text as structured data in JSON format. This structured data is then sent to the results display unit on the terminal via a secure communication channel. The notification unit monitors an event-driven message queue running on the server side. When new decision result data is added to the queue, the notification unit sends a push notification to the terminal using the WebSocket protocol. This system executes the process from decision result generation to notification completely asynchronously and in real-time. This system minimizes user waiting time. It displays the reasons for unsuitability along with specific numerical evidence. It clarifies the user's next action. For example, it displays additional warranty requirements or specific modification instructions on the screen. This system achieves the technical benefits of reducing rework in business processes and increasing overall throughput. In use cases such as occupancy verification in real estate leasing and reissue verification of credit cards, this system helps to provide rapid feedback on verification results and improve customer satisfaction.

[0064] Step 5: Automatic generation and visualization of enterprise maps The mapping department automatically categorizes companies based on various attributes, including the judgment and risk assessment results from the judgment department, the company's region, industry, transaction size, risk level, alliance partner attributes, business plan content, administrative guidance records, intellectual property holdings, employee size, capital size, and years of establishment. It generates company maps using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, and timelines. The results are displayed on the terminal, allowing department heads to refer to them in real time. For example, it can automatically generate company maps from various entry points, such as IT companies in the Kanto region with annual revenue exceeding 1 billion yen and low risk levels, companies with past administrative guidance records but now deemed to have improved, growing companies with substantial intellectual property holdings, and startups less than 3 years old. Specifically, the mapping department represents each company as data points in a multi-dimensional feature space. It combines each company's attribute data and judgment scores to generate a 256-dimensional feature vector. The mapping department performs principal component analysis or t-distributed random neighborhood embedding algorithms. This mapping unit reduces and compresses 256-dimensional feature vectors into low-dimensional coordinate data of 2 or 3 dimensions. Based on the compressed coordinate data, it calculates the similarity between enterprises using Euclidean distance. It then applies a k-means clustering algorithm to automatically divide the enterprise group into multiple clusters. The unit calculates the centroid coordinates of each cluster and the list of enterprises it contains. It generates directed graph data with transaction or capital relationships between enterprises as edges and enterprises as nodes. A graph rendering algorithm based on a mechanics model is used to calculate the optimal configuration coordinates of the nodes. These calculation results are converted into vector graphics data. For example, it generates vertex array data in SVG format or for WebGL. This graphics data is then sent to the terminal. The terminal's result display unit utilizes GPU hardware acceleration to render the graphics data at high speed. This system transforms high-dimensional and complex enterprise data groups into low-dimensional spatial configurations that can be intuitively processed by the human visual cortex. This system facilitates the discovery of hidden correlations or cluster structures between data. Compared to previous analytical methods that read massive amounts of text data, this system significantly reduces the cognitive load of information processing. This system strongly supports strategic decision-making in use cases such as supply chain vulnerability analysis, understanding the distribution of competitors in specific regions, and identifying target companies for expanding into new markets.

[0065] Step 6: Continuously improve judgment accuracy through AI learning The AI ​​Learning Management Department continuously learns from the judgment results made by the Judgment Department in the past, new information acquired from external sources, the occurrence of disputes after actual transactions, the occurrence of administrative guidance, client bankruptcy or mergers, growth performance, etc., to improve the judgment accuracy of generative AI. For example, if a company judged as unsuitable in the past causes problems later, this case is imported as learning data and reflected in future judgment criteria. In the event of new legal amendments or changes in industry guidelines, the AI ​​Learning Management Department also automatically acquires information and reflects it in the judgment department's algorithm. Specifically, this AI Learning Management Department manages a time-series database built in the server's storage area. This AI Learning Management Department accumulates the past input tensors, output scores, and judgment labels output by the judgment department as historical data. This AI Learning Management Department periodically acquires the post-event status of target companies through external linkage departments. For example, this AI Learning Management Department acquires binary data such as bankruptcy indicators, administrative guidance indicators, and debt default indicators. This AI Learning Management Department combines the acquired post-event status as correct answer labels into the historical data and automatically generates a training dataset for supervised learning. This AI Learning Management Department detects that the size of the training dataset has reached the predetermined batch size. This AI learning management system initiates a relearning process for a model using the backpropagation algorithm in a background thread. It calculates the cross-entropy error between the predicted score and the correct answer label. It then uses gradient descent to incrementally update the weight parameters of each layer of the neural network. New text data related to legal amendments or guideline changes is acquired. This text data is added to the corpus, continuously performing pre-training of a large-scale language model. Unlike static rule-based systems, this system autonomously adapts to environmental changes or new risk patterns. It delivers the technical benefits of improving model generalization performance over time and continuously reducing prediction errors on unknown data. The system automatically performs accuracy evaluation tests using a validation dataset before and after model weight updates. It only updates production model parameters via hot-swapping when accuracy improves. The system always provides up-to-date inference capabilities while maintaining zero downtime. This system offers indispensable adaptability in use cases such as compliance inspection systems dealing with frequent changes in financial regulations and security monitoring systems detecting new cyberattack techniques.

[0066] (specific use case) For example, when the sales department reviews a legal entity whose main business is selling information and communication equipment as a potential new customer, the user interface of the terminal is used first to input or upload the company name, representative name, location, date of establishment, capital, number of employees, business scope, financial statements for the past three periods (balance sheet, profit and loss statement, cash flow statement), all historical documents (PDF file), regulations regarding the personal information protection system, a list of major customers, records of administrative guidance, the URL of the company website, alliance information, business plan (sales targets, investment plan, growth strategy, risk analysis), and regional information. The input support department automatically suggests past registered company names or similar names when the company name is entered, automatically completes the address based on the postal code when the location is entered, automatically extracts numerical data from the PDF when uploading financial statements, performs OCR processing on the images of all historical documents, and automatically generates a checklist for the personal information protection system. Specifically, while the user is inputting data into the input form on the terminal, this input support department executes multiple data preprocessing pipelines in parallel in the background. This input support unit normalizes the character encoding of strings in the text field inputting company names. It then uses the normalized strings as queries and performs a fuzzy search based on Levenstein distance on a distributed cache server in memory. The unit extracts a list of company names with similarity scores exceeding a threshold in milliseconds and dynamically renders them as suggested candidates on the screen. It applies a layout parsing algorithm to uploaded PDF financial statements. It calculates border information and relative coordinates of text blocks within the document. Based on the calculated coordinates, it reconstructs the table structure and expands the values ​​of each cell as floating-point multidimensional arrays in memory. It performs image preprocessing, including denoising and contrast adjustment, on the image data of all historical certificates. The preprocessed image tensor is input into a convolutional neural network, which outputs bounding boxes for character units and probability distributions for character categories. The unit connects the character categories with the highest probabilities to generate text data. Finally, it uses natural language processing to extract specific keywords from the extracted text data. For example, it extracts strings such as personal information or management personnel. This input support department automatically generates checklist data in JSON format based on the extracted results. The system completes the conversion from unstructured data to structured data at the edge. This system offers the technical benefits of saving communication bandwidth to the server and reducing overall system processing latency. In use cases such as insurance companies digitizing and reviewing large amounts of paper media materials, or purchasing departments in manufacturing industries processing diverse formats of documents submitted by numerous suppliers, this system eliminates bottlenecks in input operations.

[0067] In addition to information received from terminals, the Information Acquisition Department automatically collects relevant information from the company's existing customer database, legal department's contract database, accounting department's financial database, sales department's transaction history database, and audit department's audit record database. For example, it extracts past registrations under the same company name or representative name, past contract details, past transaction performance, and audit findings. The External Liaison Department obtains credit ratings and financial scores from credit investigation agency databases, administrative guidance records and law violation records from publicly available information from administrative agencies, automatically acquires reputation, word-of-mouth, and news reports from corporate information websites and news websites on the internet, analyzes the frequency of company names and co-occurrence rates with negative terms from social networking sites, obtains membership qualifications and industry reputation from industry associations, and acquires the ownership and application status of intellectual property rights from patent information databases. Specifically, this Information Acquisition Department orchestrates multiple data acquisition workers based on a microservice architecture. Using the company identifier received from the terminal as the key, this Information Acquisition Department issues asynchronous RPC calls to various databases on the company's internal network. This information acquisition department merges XML or JSON records returned from various databases into a single integrated data frame according to a common data pattern. This external collaboration department sends secure requests using OAuth authentication to external Web APIs. This external collaboration department acquires structured data such as credit ratings or financial scores and stores it in the corresponding columns of the integrated data frame. This external collaboration department conducts distributed web crawling for unstructured data sources such as news websites or social networking sites. This external collaboration department parses the DOM tree from HTML documents and extracts the main text. This external collaboration department inputs the extracted text into a natural language processing pipeline. This external collaboration department tokenizes the text, removes stop words, and generates word frequency vectors. This external collaboration department calculates the inner product of this frequency vector with a negative vocabulary dictionary. This external collaboration department normalizes the inner product value and calculates a co-occurrence score in the range of 0 to 1. This external collaboration department converts a list of patent classification codes obtained from a patent information database into a multi-dimensional vector using one-hot encoding. This system automatically constructs a unified tensor representation that can be processed by machine learning models from heterogeneous data sources scattered inside and outside the company. This system offers a significant technological boost by improving the comprehensiveness of data collection and eliminating information silos. In use cases such as corporate valuation in investment banking and the selection of new business partners in general trading companies, this system can achieve detailed risk profiling based on multi-faceted information.

[0068] The judgment department inputs all information collected by the information acquisition department into the generative AI. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profits, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, the increase or decrease of fixed assets, the content of investment activities, etc., and compares them with industry averages or past customer data to score the soundness of finances. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extract the legitimacy of legal person status, the history of changes in representatives, the content recorded in the purpose column, the history of increases or decreases in capital, the history of changes in location, records of dissolution or liquidation, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection systems, the generative AI parses checklists or regulatory documents, automatically determines their suitability with laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, whether there are countermeasures against information leakage, and the status of obtaining third-party certifications, etc. Regarding past business performance and administrative guidance records, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, content of administrative guidance, existence of improvement reports, implementation status of countermeasures to prevent recurrence, etc., to assess the risk level. Regarding the content recorded on the company's website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, differences from past records, accuracy of product or service descriptions, whether there are any violations of laws and regulations, and whether there is any infringement of third-party rights. Regarding alliance information, generative AI analyzes the credit information, past transaction performance, reputation in the industry, appropriateness of cooperation content, contract content, etc. of partner companies to assess risks and synergies. Regarding business plans, generative AI automatically summarizes the content of business plans and analyzes the appropriateness of sales targets, the realism of investment plans, the feasibility of growth strategies, the comprehensiveness of risk analysis, comparison with similar past cases, and consistency with industry trends. Specifically, this judgment department segments the integrated data frames according to modality and inputs them into the corresponding neural network models. This decision-making unit normalizes the numerical array of financial information and inputs it into a Long Short-Term Memory (LSTM) network to capture time-series dependencies. The final hidden state vector of the LSM network is then fed into a fully connected layer, outputting a financial soundness score as a continuous value from 0 to 100. Text data extracted from all historical documents is then input into a pre-trained language model. Attention weights are calculated for each word tag within the text, extracting feature vectors representing specific events such as changes in location or other events. The cosine similarity between this feature vector and a risk pattern vector generated from a database of past violations is calculated. If the similarity exceeds a threshold, a risk flag is set to 1.This decision-making unit segments the text data of the prescribed document into sentences, converting each sentence into a sentence embedding vector. It calculates a pairwise similarity matrix between the vectors of each clause in the legal guidelines. The unit performs max pooling of the similarity matrix to calculate a suitability score of 0 to 1, representing the comprehensiveness of the prescribed content. This unit applies a sentiment analysis model to the text data of news reports or websites. It outputs a probability distribution of positive, negative, and neutral values ​​for the entire text. If the negative probability exceeds a predetermined value, the unit increases the risk level by one stage. This unit inputs the text data of a business plan into a summarization model, extracting key sentences to generate a summary text vector. It maps the summary text vector to the vector space of past successes and failures, using the k-nearest neighbors method to estimate the success probability. The system combines the scores, labels, and probability values ​​output from each model into a single feature vector. This feature vector is then input into a gradient boosting decision tree algorithm to calculate the final comprehensive evaluation score. This system transforms diverse unstructured data into high-dimensional mathematical representations and evaluates them through quantitative computation. This system offers the technical advantage of eliminating evaluation biases caused by human subjectivity, ensuring the objectivity and reproducibility of reviews. In use cases such as the review of franchise stores or the pre-review of crowdfunding projects, this system can process a large number of cases quickly and with high accuracy.

[0069] For example, specific examples of prompts that generative AI can read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal status and past risk factors from the contents of all the following historical certificates and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following company websites and determine whether there are any false records or exaggerated advertisements," "Summarize the following business plan and compare it with similar past cases to analyze growth and risks," and "Extract negative topics or online criticism risks about the company name from the following SNS post data and calculate the risk score," etc. Specifically, this system dynamically constructs these natural language-based prompts in memory as system prompts interpretable by a large-scale language model. The system uses a tokenizer to convert the prompt string into an array of integer indices. This index array is input into the embedding layer, where each token is converted into a multi-dimensional dense vector. The system also vectorizes the input data group as objects. For example, this system vectorizes numerical arrays from financial statements or text data from SNS submissions. It combines the vector sequence of prompts and the vector sequence of input data to generate a single input tensor. For instance, this system receives the text string "The product is about to break down" and the target company name as input for an SNS submission. This system tokenizes these texts and appends them to the end of the tokenized sequence of the prompt. The combined input tensor is then fed into the self-attention mechanism of the Transformer model. This model calculates a high attention weight between the prompt's indication and the input data. The model outputs a floating-point number in the range of 0.0 to 1.0 representing the risk of online criticism, along with the extracted text used as the basis. For example, this model outputs a score of 0.85 and the string "The product is about to break down." The system then passes the output floating-point number to the subsequent thresholding module. This system transforms flexible instructions based on natural language into deterministic matrix operations in a high-dimensional space. This system reduces the cost of developing and learning individual, dedicated models for each task. This system offers the technical advantage of performing diverse analytical tasks in parallel using a single basic model. It provides high versatility and processing efficiency in use cases such as automatic customer voice classification in customer support and quantitative measurement of brand image in marketing departments.

[0070] The judgment department scores each item, performs a comprehensive suitability assessment, and generates detailed scoring results or supporting data, AI-based summary comments, comparisons with similar past cases, etc., which are then sent to the results display unit on the terminal. The notification department notifies the user in real time of the judgment results or reasons for unsuitability, risk alerts, requests for additional information, and feedback on AI learning results. For example, if financial soundness falls below a certain standard, there is a significant past record of administrative guidance, false information is confirmed on the website, the personal information protection system is insufficient, risks are confirmed by the alliance party, or the business plan is unrealistic, the condition is deemed unsuitable, and the reasons are automatically generated and notified to the terminal. Specifically, this judgment department combines the individual score groups output from each neural network model as one-dimensional feature vectors in memory. This judgment department multiplies this feature vector by a pre-learned weight matrix. This judgment department calculates a weighted average score that incorporates the importance of each item. This judgment department inputs the calculated weighted average score into a Sigmoid function and outputs a comprehensive suitability probability in the range of 0 to 1. If the overall suitability probability is less than a pre-set threshold, this decision-making unit sets an unsuitability flag in a register in memory. For example, this decision-making unit sets the threshold to 0.7. When the unsuitability flag is set, this decision-making unit identifies the element in a specific feature vector that deviates most from the threshold. For example, this decision-making unit identifies the financial soundness score as a primary factor. This decision-making unit inputs the identifier of the identified element into a natural language generation model. This decision-making unit dynamically generates a text string explaining the reasons for unsuitability. For example, this decision-making unit generates a string explaining that the financial soundness score is 42, which is below the benchmark value of 50. This decision-making unit serializes the score, flag, and generated text string as a JSON object. This decision-making unit forwards this JSON object to the notification unit. The notification unit parses the received JSON object. The notification unit asynchronously pushes this data via a persistent TCP connection established with the user's terminal. This system automatically converts the results of complex multivariate analysis into structured data and natural language text that humans can immediately understand. This system eliminates the cognitive load of the reviewer interpreting massive amounts of raw data. This system has the technical effect of completing the feedback loop of review results in milliseconds. This system enables rigorous risk management without compromising user experience in use cases such as real-time review on online financing platforms or automatic background checks during platform registration for gig workers.

[0071] The mapping department automatically categorizes companies based on various attributes, including the judgment or risk assessment results made by the judgment department, the company's region, industry, transaction size, risk level, alliance partner attributes, business plan content, administrative guidance records, intellectual property ownership status, number of employees, capital size, and years of establishment. It then generates company maps using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, and timelines. The results are displayed on the terminal, allowing department heads to refer to them in real time. For example, it can automatically generate company maps based on various entry points, such as IT companies in the Kanto region with annual revenue exceeding 1 billion yen and low risk levels, companies with past administrative guidance records but now deemed to have improved, growing companies with substantial intellectual property holdings, and startups less than 3 years old. Specifically, the mapping department loads the multi-dimensional attribute data of each company extracted from the database into data frames in memory. The mapping department also parses the filtering conditions sent from the user terminal. For example, this mapping department analyzes conditions such as the region being Kanto, the industry being IT, annual revenue exceeding 1 billion, and a risk score less than 0.3. This mapping department applies logical multiplication to the data frame to quickly extract a subset of enterprise records that meet the conditions. Using the extracted subset data, this mapping department dynamically generates a data structure corresponding to the specified visualization method. For example, in the case of generating a network graph, this mapping department constructs an adjacency matrix with enterprises as nodes and transaction or capital relationships between enterprises as edges. This mapping department applies an eigenvector centrality algorithm to the adjacency matrix to calculate the importance score of each node. This mapping department determines the depiction size of the nodes based on the importance score. This mapping department assigns RGB color codes to the nodes according to the risk level. This mapping department generates graph data in JSON format containing these depiction parameters and sends it to the terminal. The results display unit on the terminal side uses the WebGL library to render the received graph data onto a canvas. Based on user interactions, this system dynamically recalculates and redraws tens of thousands of enterprise data entries in milliseconds. This system visually reveals hidden inter-firm relationships or risk biases in specific market segments that are difficult to detect in static reports. It delivers a significant technological boost to the efficiency of data exploration. In use cases such as monitoring macro-systemic risks for financial authorities or optimizing global supply chains for multinational corporations, this system provides a data-driven decision-making foundation.

[0072] The AI ​​Learning Management Department continuously learns from past judgments made by the Judgment Department, newly acquired external information, the occurrence of disputes after actual transactions, the occurrence of administrative guidance, client bankruptcies or mergers, and growth performance to improve the accuracy of generative AI judgments. For example, if a company previously deemed unsuitable later causes problems, this case is imported as learning data and reflected in future judgment benchmarks. In the event of new legal amendments or changes in industry guidelines, the AI ​​Learning Management Department also automatically acquires information and reflects it in the Judgment Department's algorithm. Specifically, the AI ​​Learning Management Department uses daily batch processing to ingest the latest transaction data and external event logs into a data lake built within the system. The AI ​​Learning Management Department combines the input feature vectors from past judgment time points with subsequent actual events to construct a new teacher dataset. For example, the AI ​​Learning Management Department incorporates binary flags indicating the presence or absence of debt defaults within one year of the transaction's commencement. The AI ​​Learning Management Department uses the constructed teacher dataset to fine-tune the neural network model. The AI ​​Learning Management Department uses stochastic gradient descent for error backpropagation to update the model's weight matrix. This AI learning management department dynamically increases the penalty weight in the loss function for false positives that were previously deemed suitable but resulted in problems. This AI learning management department adjusts the direction of learning to prevent the model from repeating the same errors. This AI learning management department periodically polls external official gazette databases or the Ministry of Justice's API. This AI learning management department acquires text data of new legal amendments. This AI learning management department tokenizes the acquired text data and utilizes it as an additional pre-training corpus for large-scale language models. This system eliminates the need for manual rewriting of rule-based conditional expressions by human engineers in previous maintenance work. This system establishes a data-driven autonomous model update pipeline. This system has the technical effect of continuously and automatically adapting the system's predictive accuracy to changing economic environments or new violations over time. This system is extremely effective in use cases such as credit card misuse detection systems or spam filtering systems that require continuous learning of the latest patterns in hostile environments.

[0073] (Application Example 1) The specific processing flow corresponding to Use Case 1 will be described below. The various parts of the system described below are implemented by the data processing device 12 and the intelligent device 14. Furthermore, the data processing device 12 is referred to as the "server" and the intelligent device 14 is referred to as the "terminal".

[0074] (Implementation methods for carrying out the present invention) As an embodiment of the present invention, the franchise store review and management system in the electronic settlement service is constructed by connecting a server and a terminal via a network. Multiple functional modules are installed on the server side, including an information acquisition unit, a judgment unit, a mapping unit, a database management unit, an external linkage unit, an AI learning management unit, a log management unit, and a security management unit. On the terminal side, a user interface unit, an input support unit, a result display unit, a notification unit, a file upload unit, a download unit, and a user authentication unit are installed. Specifically, this franchise store review and management system adopts a distributed system architecture built on a cloud computing environment. The system deploys each functional module on the server side as an independent container. The system uses container orchestration tools to dynamically expand the resource allocation of each module. For modules with high computational loads, such as the information acquisition unit or the judgment unit, the system prioritizes the allocation of GPU instances. The system uses high-speed binary communication via the gRPC protocol to achieve communication between modules. The system installs each part on the terminal side as a single-page application running on a web browser. The system constructs a virtual DOM in the terminal-side memory to perform user interface rendering and updates with minimal computational cost. In the User Authentication Department, this system performs mutual authentication using digital certificates based on public-key cryptography. The system encrypts data packets along the communication path using the TLS protocol. In the Security Management Department, all API requests undergo token-based authorization checks. In the Log Management Department, all state migrations and data access histories within the system are stored in an immutable, blockchain-based distributed ledger. This system eliminates performance bottlenecks stemming from monolithic system design. This system delivers high scalability and availability, ensuring the processing of millions of pieces of data from franchisees. This system provides a robust and high-speed infrastructure for use cases such as large-scale e-commerce platforms deployed globally, or payment gateway services integrating multiple mobile payment applications.

[0075] First, end-users (e.g., franchisee review managers, risk management managers, sales managers, legal managers, accounting managers, system integration managers, audit managers, etc.) input or upload the following information through the user interface: franchisee candidate's company name, representative name, location, date of establishment, capital, number of employees, business scope, financial statements for the past three periods (balance sheet, profit and loss statement, cash flow statement), all historical documentation (PDF or image files), regulations regarding personal information protection (e.g., privacy policy, information management regulations, employee training records), pledges to exclude antisocial forces, past records of illegal transactions or administrative guidance, company website URL, industry classification, and technical information used for settlement system integration (API specifications, security certification information, system architecture diagram, communication protocol description, encryption method description, sandbox test results, etc.). For example, financial statements are uploaded as Excel files or PDFs, and all historical documentation is attached as image files. Information regarding personal information protection is entered as a checklist or document file, and the pledge to exclude antisocial forces is submitted as an electronically signed PDF. Regarding technical information, upload API specifications or system architecture diagrams, detailed communication protocols, encryption method descriptions, and sandbox test results reports. Specifically, this user interface unit renders dynamic input forms on the terminal's display. This user interface unit temporarily buffers user-inputted text data or uploaded binary files in the terminal's local memory. This file upload unit performs MIME type verification and malware scanning on the buffered file group in the background. This file upload unit asynchronously sends the secure file as multi-part form data to the server. Upon receiving a PDF file with an electronic signature, this system performs cryptographic verification of the signature using a public key infrastructure. This system uses a hash function to verify the integrity of the file and deterministically determines whether it has been tampered with. This system parses API specifications uploaded as technical information. For example, this system parses YAML files in OpenAPI format. This system expands the endpoint URL, request parameter data types, and response patterns as in-memory data structures. This system applies object detection algorithms to image files of the system architecture diagram. This system identifies components such as firewalls, load balancers, and databases, and extracts the network topology as graph data. This system automatically preprocesses unstructured and semi-structured data provided in various formats, converting them into a unified internal representation that can be processed by subsequent machine learning models. This system effectively eliminates human error during data input and ensures data quality in the early stages of the review process. In use cases such as reviewing new token listings on cryptocurrency exchanges or verifying connections from third-party providers utilizing open banking APIs, this system efficiently collects and verifies complex technical requirements.

[0076] The input support department automatically suggests past registered company names or similar names when a company name is entered, provides a list of past representatives when a representative name is entered, automatically completes the address based on the postal code when a location is entered, automatically extracts numerical data from PDFs or Excels when financial statements are uploaded, performs OCR processing on images of all historical certificates, automatically generates a checklist for personal information protection systems, and performs format conversion of API specifications or automatic parsing of communication protocols. The file upload department automatically determines the format of uploaded files, extracts text from PDFs, and performs OCR processing on images as needed to achieve data standardization. The user authentication department has security functions such as single sign-on or multi-factor authentication, controls access permissions by department, and leaves an audit trail. Specifically, this input support department sends an asynchronous AJAX request to the server with each keystroke by the user. This input support department performs high-speed string matching using N-gram indexes on the server-side in-memory database. This input support department replies to the terminal with a list of candidates with the highest similarity scores in milliseconds and displays it as a drop-down menu. When an Excel file is uploaded, this input support department parses the file header and decodes the OpenXML format. This input support unit extracts numerical values ​​such as sales revenue or operating profit based on specific cell addresses or regular expression patterns and converts them into floating-point array data. This input support unit inputs text data from API specifications into the syntax parser. This input support unit automatically identifies protocol types such as REST or GraphQL and converts them into an internal standardized JSON schema. This user authentication unit uses SAML or OAuth 2.0 protocols to interact with external identity providers. This user authentication unit issues a JWT containing access permission information upon successful user authentication. This user authentication unit verifies the signature of the JWT included in the HTTP header in all subsequent API requests. This user authentication unit appends the request timestamp, IP address, and executed operation content to an audit log database with a hash chain structure. This system not only reduces the physical hassle of data input but also enforces normalization of input data format at the system level. This system dramatically reduces the incidence of parsing errors or exception handling in subsequent data processing pipelines. This system ensures both data accuracy and security in use cases such as data input support for electronic medical record systems in medical institutions or automated processing of various application forms in administrative agencies.

[0077] The information entered from the terminal is sent to the information acquisition department on the server side. In addition to the information received from the terminal, the information acquisition department automatically collects relevant information from the company's existing franchisee database, past review history, transaction performance database, legal department's contract database, accounting department's financial database, system linkage history database, and audit record database. For example, to verify existing franchisees, a search is performed using keys such as company name, representative name, location, and establishment date to automatically retrieve whether the same or similar franchisees have been registered in the past. From past review history and transaction performance, it also retrieves records of past illegal transactions or disputes, details of contract terminations or transaction cessations, and audit findings. Furthermore, through external coordination, the system automatically obtains credit ratings and financial scores from credit investigation agency databases, administrative guidance records and law violation records from publicly available information from administrative agencies, reputation and word-of-mouth from corporate information websites and news websites on the internet, the frequency of company names and their co-occurrence with negative terms from social networking sites (SNS), membership and industry reputation from industry associations, the status of intellectual property rights and application status from patent information databases, and past system event information from cybersecurity vulnerability information databases. For example, the system automatically collects information such as the risk of online criticism of company names on SNS, past administrative guidance records, expulsion history from industry associations, the number of patent applications and the content of held patents, cyberattack victimization history, and the existence of vulnerability reports. Specifically, this information acquisition department initiates multiple data collection processes via a distributed messaging system. This information acquisition department performs normalization processing on the input company name and representative name strings to absorb textual fluctuations. For example, this information acquisition department normalizes differences between full-width and half-width characters and differences in the representation of joint-stock companies. This information acquisition unit hashes the normalized string and performs a constant-time search on the existing database index. It extracts time-series values ​​such as chargeback rates and settlement cancellation rates from a historical transaction performance database, expanding them as multi-dimensional time-series tensors in memory. This external linkage unit sends queries to a cybersecurity-related vulnerability information database, using the target company's domain name and the identifier of the software stack it's using. It retrieves a list of returned vulnerability scores, calculates the maximum and average values, and stores them in a feature vector. This external linkage unit connects to the SNS streaming API to acquire submission text containing the target company's name in real time. It inputs the acquired text into a natural language processing model to calculate a sentiment polarity score. For example, it calculates a continuous value from -1.0 to +1.0. This external linkage unit calculates a moving average of negative scores over a certain period and sets a risk flag for online criticism if the score exceeds a threshold.This system integrates structured data from within the company and unstructured data from outside, using a common identifier as the key, into a single high-dimensional data space. It possesses the technical effect of amplifying subtle risk signals that might be missed in a single data source through cross-validation from multiple data sources, making them detectable. This system functions as a comprehensive threat intelligence foundation in use cases such as anti-money laundering checks in cryptocurrency exchanges and security audits of new tenants in cloud service providers.

[0078] The vast amount of information collected by the Information Acquisition Department is sent to the Judgment Department on the server side. The Judgment Department uses generative AI to analyze and judge various types of information. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculating sales, profits, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, increases and decreases in fixed assets, the content of investment activities, the shift of short-term and long-term loans, increases and decreases in inventory, and whether there are any tax arrears, etc., and compares them with industry averages or past franchise store data to score the soundness of finances. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extracting the legitimacy of legal person status, the history of changes in representatives, the content recorded in the purpose column, the history of increases and decreases in capital, the history of changes in location, records of dissolution or liquidation, mergers, divisions, business transfers, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection mechanisms, generative AI analyzes checklists or regulatory documents to automatically determine their suitability to laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, the existence of countermeasures against information leaks, the acquisition status of third-party certifications (information security certifications, privacy marks (registered trademarks), etc.), whether there is entrusted management, and the completeness of personal information management ledgers. Regarding the exclusion of antisocial forces, the presence of risk is determined by comparing the content of pledges or external databases, such as comparing with antisocial force lists and verifying past administrative penalty histories. Regarding past records of illegal transactions and administrative guidance, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, the content of administrative guidance, whether there are improvement reports, the implementation status of countermeasures to prevent recurrence, the frequency and severity of administrative guidance, etc., to assess the risk level. Regarding the content recorded on the company website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, differences from past recorded content, the accuracy of product and service descriptions, whether there are any legal violations, whether there are any infringements of third-party rights, and whether there are any false representations or violations of the gift representation law. Regarding industry risks, AI determines whether an industry is high-risk by referring to the overall industry's violation rate and administrative guidance rate, past franchise dispute rate, industry group expulsion rate, industry-specific regulatory risks, and social credit. Regarding the suitability of the settlement system integration, technical information and API specifications are analyzed to automatically determine compatibility with existing settlement systems, compliance with security requirements, encryption strength, standard compliance with communication protocols, sandbox testing pass rate, past system failure history, and the presence of external vulnerability reports. For example, by automatically comparing the content recorded in the API specifications with the specifications of the settlement service side, AI automatically determines compatibility and the presence of security vulnerabilities, and analyzes sandbox test logs to detect any illegal communication or abnormal responses.Specifically, this decision-making unit inputs a variety of collected data into a multimodal neural network based on the Transformer architecture. It inputs a numerical array of financial information into a multilayer perceptron to generate a 64-dimensional financial feature vector. It inputs textual data from all historical certificates and regulatory documents into a large language model to generate a 256-dimensional textual feature vector. It parses the structured text of API specifications into a syntax tree and inputs it into a graph neural network to generate a 128-dimensional technical feature vector. It inputs communication logs from sandbox testing into a one-dimensional convolutional neural network to extract feature vectors representing abnormal communication patterns. For example, it inputs a time-series array of HTTP status codes and the distribution of response times. This decision-making unit combines the feature vectors obtained from these various modalities to construct a single high-dimensional integration tensor. This integration tensor is then input into a final classification layer with a self-attention mechanism. This classification layer calculates the complex nonlinear interactions between the features. For example, this classification layer learns weights so that when both the characteristics of low financial soundness and outdated API encryption are present simultaneously, the output risk score is exponentially higher than when either characteristic exists alone. This decision unit, as the final output, represents a probability value of 0.0 to 1.0 for eligibility as a franchisee, along with individual scores for each risk category. This system formalizes the comprehensive risk assessment across multiple domains, which is intuitively performed by human reviewers, into tensor operations in high-dimensional mathematical space. This system dramatically improves the accuracy and consistency of reviews. This system delivers the technical benefit of reducing review time from weeks to minutes. In use cases such as API-linked reviews of new services in fintech companies and security audits of participating institutions in medical data sharing platforms, this system automates judgments requiring high levels of expertise, supporting the construction of a secure ecosystem.

[0079] The judgment department scores each item individually, performs a comprehensive suitability assessment, generates detailed scoring results along with supporting data, AI-generated summary comments, comparisons with similar past cases, etc., and sends them to the results display unit on the terminal side. For example, if financial soundness falls below a certain standard, there is a significant past record of administrative guidance, false information is found on the website, personal information protection is inadequate, antisocial forces are detected, industry risk is high, there are technical issues with the settlement system, the product fails sandbox testing, or there are external vulnerability reports, the product is deemed unsuitable, and the reasons are automatically generated and notified to the terminal side. Conversely, if all items meet the standards, the product is deemed suitable for franchise registration, and registration is processed. Specifically, the judgment department inputs the overall suitability probability and the individual scores for each category into a rule engine deployed in memory. This rule engine follows predefined decision tree logic, evaluating whether each score meets the prescribed threshold conditions through logical operations. If the judgment department detects a fatal risk in any category, it performs a short-circuit assessment to immediately determine the product as unsuitable. For example, this decision-making unit performs a short-circuit assessment if the antisocial forces risk score is above 0.9 or if the fatal error flag in the sandbox test is 1. If deemed unsuitable, this decision-making unit specifies the elements of the feature vector that trigger it. This decision-making unit inputs the identifiers of the specified elements into a natural language generation model to generate text explaining the specific reasons for unsuitability. For example, this decision-making unit generates the string "Because the API authentication method uses a deprecated algorithm, it does not meet security standards." If the scores of all items pass the threshold, this decision-making unit makes a comprehensive judgment as suitable. Upon obtaining a suitable judgment, this decision-making unit automatically issues an INSERT query to the database management department targeting the franchisee's master table. This decision-making unit asynchronously sends a request to the settlement gateway API to issue a production environment API key for new franchisees. This system automates the entire process from the generation of the judgment result to subsequent system registration processing without human intervention. This system achieves completely direct processing of business processes. This system has the technical effect of minimizing operating costs and the time required for new franchisee onboarding. This system provides an indispensable automated foundation to support business scale in use cases such as rapidly growing electronic payment platforms and review processes of app stores with a large number of third-party developers.

[0080] The registered franchise information is automatically categorized by the server-side mapping department based on multiple attributes, including industry, region, transaction scale, risk level, settlement system integration status, administrative guidance record, intellectual property ownership, number of employees, capital size, years of establishment, cybersecurity risk level, and past system failure history. For example, it automatically generates franchise maps from various perspectives, such as high-risk groups in the catering industry, groups with completed settlement system integration in the IT industry, groups with past administrative guidance records but now deemed improved, growing companies with substantial intellectual property rights, startup franchises less than 3 years old, high-risk franchises, and franchises with a history of causing multiple system failures. The mapping department displays this information using various visualization techniques, including graph structures, maps, heatmaps, network diagrams, tree diagrams, timelines, risk heatmaps, and cybersecurity risk maps. The results are displayed on the terminal side, allowing department heads to refer to the information in real time. For example, risk management managers can use heatmaps to identify the distribution of franchisees in high-risk industries, sales managers can use maps to identify the distribution of franchisees in newly developed areas, legal managers can use timelines to track the time series of franchisees with administrative guidance records, and system managers can use risk maps to identify franchisees with high cybersecurity risks. Specifically, this mapping department uses high-dimensional vector space clustering algorithms. It converts franchisee information into multi-dimensional feature vectors. It converts industry information into binary vector data using one-hot encoding. It performs logarithmic transformation and normalization on continuous data such as capital and number of employees. It combines these features to generate a single franchisee tensor. It uses principal component analysis or t-distributed random neighborhood embedding to reduce the dimensionality of the generated franchisee tensor to two- or three-dimensional spatial coordinates. Based on the reduced coordinate data, it calculates the similarity between franchisees using Euclidean distance or cosine similarity. Based on the calculated similarity, it performs k-means or hierarchical clustering. This mapping department constructs the clustering results as graph structure data in memory. This mapping department defines nodes as individual franchisees. This mapping department defines edges as transaction relationships or similarities between franchisees. This mapping department performs calculations to make the node depiction size proportional to the capital scale. This mapping department changes the node depiction color from red to blue according to the network security risk level. This mapping department inputs the graph structure data into the rendering engine. This rendering engine utilizes the hardware acceleration capabilities of the parallel computing processor to perform rendering processing in real time. This system sends the rendering results to the result display unit on the terminal side via the network. This system does not rely on human visual perception processes, directly visualizing the mathematical relationships of high-dimensional data. This system thus clarifies complex risk structures that are difficult for humans to intuitively grasp. This system significantly improves the processing speed of data parsing. This system optimizes memory consumption.This system is specifically applied in the franchisee review process of financial institutions and the franchisee management of e-commerce platform operators. It enables risk management managers to instantly identify the distribution of high-risk industries. It supports sales managers in using data-driven methods to discover gaps in newly developed areas. It allows legal managers to read the cascading effects of past administrative guidance records from network diagrams. It allows system administrators to visually identify vulnerable clusters of nodes that are likely to be targets of cyberattacks. This mapping department accepts interactive operations from users. When a specific node is selected, it retrieves detailed attribute information related to that node from the database via asynchronous communication. The retrieved information is displayed as an overlay in pop-up windows. The rendering resolution is dynamically adjusted based on changes in the display magnification. This allows for seamless switching between wide-area overview displays and detailed local displays. This mapping department features animation display capabilities based on time-series data. It dynamically depicts the growth process and risk changes of franchisees from the past to the present along a timeline. This allows the extraction of signs of temporal changes that are easily overlooked in static data analysis. This system centrally performs these highly detailed visualization processes on the server side. This system significantly reduces the computational load on the terminal side. This system ensures comfortable operation even from low-spec terminals.

[0081] Furthermore, the server-side AI learning management department continuously learns from past judgments made by the judgment department, newly acquired external information, the occurrence of disputes after actual transactions, the occurrence of administrative guidance, the bankruptcy or merger of franchise stores, growth performance, the occurrence of cyberattacks, and the occurrence of system failures, thereby improving the accuracy of generative AI judgments. For example, if a franchise store previously deemed unsuitable later causes problems, this case is imported as learning data and reflected in future judgment criteria. In the event of new legal amendments or changes in industry guidelines, the AI ​​learning management department also automatically acquires information and reflects it in the judgment department's algorithm. The AI ​​learning management department verifies past judgment results and actual subsequent risk occurrences, continuously optimizing the judgment logic. For example, it learns from past system linkage information and vulnerability reports of franchise stores that have experienced cyberattacks and reflects this in future cyber risk judgments. Specifically, this AI learning management department builds a continuous learning pipeline on the server. This AI learning management department automatically generates teacher data that matches past judgment results with actual subsequent data. This AI Learning Management Department uses post-event data, acquiring dispute occurrence markers and administrative guidance markers within a certain period from the start of the transaction as binary data. This AI Learning Management Department then converts this teacher data into a dataset for micro-calls of a large language model. As input data, this AI Learning Management Department constructs a text sequence containing franchisee attribute information and past system interaction logs. As a specific example of input data, this AI Learning Management Department accepts the numerical data of a capital of ten million yen and the text data of past reports of cross-site scripting vulnerabilities. As output data, this AI Learning Management Department defines a continuous value from zero to one representing the probability of risk occurrence, and a text tag column explaining the causes of the risk. As a specific example of output data, this AI Learning Management Department outputs the numerical value of a network attack victim probability of 0.85 and the explanatory text of insufficient encryption strength of the authentication system. This AI Learning Management Department uses the error backpropagation method to update the weight parameters of the neural network. This AI Learning Management Department uses cross-entropy error or mean squared error as the loss function. This AI Learning Management Department uses adaptive moment estimation as the optimization algorithm. This AI Learning Management Department performs gradient descent while dynamically adjusting the learning rate. To prevent overfitting, our AI Learning Management Department applies dropout or weight decay techniques. Upon receiving new legal amendments, the department updates the external knowledge base using retrieval-enhanced generation techniques. This allows the department to apply the latest decision criteria to the inference process without waiting for the model to relearn. As a subsequent process, the department deploys the updated weight parameters to the decision engine. Before deployment, the department performs accuracy evaluation of the model using a validation dataset. The department only allows the model to be applied to the production environment if the accuracy exceeds a certain threshold.This system eliminates existing review methods that rely on human rules of experience. Through a data-driven, self-evolving mechanism, it continuously improves judgment accuracy. It significantly enhances the detection rate of unknown network attack patterns and novice violations. It achieves the technical effect of simultaneously suppressing computational resource consumption and preventing model obsolescence. Specifically, it is applied in credit card company franchise management systems and e-commerce platform seller review systems. It implements dynamic risk scoring to prevent post-event risks. The AI ​​learning management department incorporates reinforcement learning algorithms. It receives reward signals based on whether an action's outcome is successful or unsuccessful. It uses policy gradient optimization to maximize rewards. It employs active learning to prioritize extracting data from boundary regions where the model hesitates during judgment. It provides the extracted data to human experts for annotation. The expert feedback is immediately incorporated into the learning data. This allows the AI ​​learning management department to efficiently improve model performance with limited data. This AI learning management department utilizes a distributed learning framework to perform learning processes in parallel across multiple computing nodes. This significantly reduces learning time, enabling model updates on a daily or hourly basis.

[0082] The Database Management Department provides unified management of all franchisee information, judgment results, scoring data, mapping data, AI learning data, user operation logs, external linkage history, and network security incident information, and saves this information as an audit trail. The Log Management Department meticulously records all operation history, data acquisition history, judgment history, mapping history, system linkage history, and network security incident history, enabling the audit and compliance departments to refer to them later. The Security Management Department implements data encryption, access control management, unauthorized access detection, data backup, recovery functions in case of failure, network attack detection, and abnormal communication detection to ensure the security and reliability of the entire system. Specifically, this Database Management Department adopts a hybrid structure combining relational and non-relational databases. The basic attributes and scoring results of franchisees, which are structured data, are stored in tables within a relational database. PDF files, image data, and AI-generated long texts, which are unstructured data, are stored in a document-oriented non-relational database. To improve search speed, the Database Management Department automatically builds B-tree indexes or hash indexes for frequently searched columns. To ensure data consistency, the Database Management Department implements distributed transaction control using a two-phase commit protocol. The Log Management Department converts all events occurring within the system into a structured log format. It assigns timestamps, user identifiers, session identifiers, operation categories, and object resource identifiers to log data. The Log Management Department asynchronously forwards generated log data to the log collection server via a message queue system. It builds a full-text search engine index on the log collection server. This enables searching for specific audit trails in milliseconds from hundreds of millions of log entries. To prevent log tampering, the Log Management Department uses hash chaining, calculating the hash value of each log entry and linking it to the hash value of the immediately preceding log entry to generate a new hash value. The Security Management Department uses Advanced Encryption Standard (AES) algorithms to encrypt all stored data. It rigorously manages encryption keys within a dedicated hardware security module. It uses Transport Layer Security (TLS) protocols to encrypt all communication data flowing over the network. Based on the concept of a zero-trust architecture, even access from the company's internal network always requires authentication and authorization. This security management department employs a machine learning-based anomaly detection model. Using this model as input data, the department acquires time-series vectors containing user login times, access source IP addresses, operation frequency, and data transfer volume. By utilizing an autoencoder neural network, the department detects abnormal access patterns that deviate from typical behavior in real time. Upon detecting anomalies, the department immediately terminates the session and sends an alert to administrators. Through these technical processes, the system minimizes the risk of information leakage.This system maintains a high level of availability and integrity. Specifically, it is used in the core systems of financial institutions with stringent compliance requirements and in the database management systems of government agencies handling confidential information. This system significantly reduces the human resource costs required for audit responses. It also enhances resilience against cyberattacks.

[0083] On the terminal side, the user interface provides comprehensive input support and auto-completion functions, file upload functions, result display functions, notification functions, download functions, and user authentication functions. For example, it displays easily understandable results such as suitability judgments and reasons for unsuitability sent from the judgment department, visualization information of franchise store maps, AI summary comments, risk alerts, requests for additional information, and network security risk alerts, and allows downloading in PDF or Excel format as needed. The notification department notifies users in real time of judgment results and requests for additional information, risk alerts, feedback on AI learning results, and emergency notifications in the event of network incidents. Specifically, this user interface is installed as a single-page application running on a web browser. This user interface adopts a component-oriented design for the front end, managing each screen element as an independent module. This user interface uses a state management library to maintain the data state of the entire application in a unified manner. This user interface utilizes a virtual document object model to quickly redraw only screen differences. This user interface thereby reduces the number of communications with the server and minimizes screen migration delays. The input support function of this user interface sends a query to the server using asynchronous communication whenever the user enters text in a text box. This user interface unit immediately displays the candidate list returned by the server as a drop-down menu. The file upload function of this user interface unit splits large files into multiple smaller chunks. This user interface unit sends the split chunks to the server in parallel. This user interface unit has a recovery function that restarts the upload of unsent chunks even in the event of a network interruption. This notification unit establishes a WebSocket connection with the server as a bidirectional communication protocol. This notification unit receives the payload of the push notification the moment the server-side risk assessment is completed. This notification unit parses the received payload and displays it as a toast notification in the lower right corner of the screen. In the case of high-urgency network event notifications, this notification unit displays a modal window covering the entire screen, forcing the user to confirm. The download function of this user interface unit receives the binary data generated by the server as a Blob object on storage. This user interface unit generates a temporary download URL from this Blob object, allowing the user to save the file to their local storage. This user authentication function utilizes an open authentication protocol and works with external identity providers. In addition to password input, this user authentication function also requires multi-factor authentication using biometric authentication or one-time passwords. Through these technical processes, this system significantly reduces the user's input burden. This system prevents erroneous input caused by human error. Through real-time information transmission, it accelerates business decision-making. Specifically, it is applied in large-scale business support systems accessed simultaneously by multiple responsible parties, and as a secure information sharing platform in remote work environments. This system delivers a technical advantage that balances intuitive operation with a high level of security.This user interface employs a responsive design, automatically optimizing its layout for various screen sizes, including personal computers, tablets, and smartphones. It adheres to accessibility standards, supporting both text-to-speech via screen readers and keyboard-only operation.

[0084] Thus, in this embodiment, the main functions such as the information acquisition department, judgment department, mapping department, AI learning management department, database management department, external linkage department, log management department, and security management department are implemented on the server side, while the functions such as the user interface department, input support department, result display department, notification department, file upload department, download department, and user authentication department are implemented on the terminal side. Generative AI, as an essential component, is incorporated into the judgment department and AI learning management department on the server side and is applied to various information processing tasks such as parsing, judgment, learning, summarizing, comparing, visualizing, and network risk assessment. For example, by combining various functions of generative AI, such as document parsing based on natural language processing, extraction of certificate content based on image recognition, risk assessment based on anomaly detection, automatic summarization of business plans based on summary generation, prompts for judgment criteria based on similar case searches, automatic capture and summarization of external information, trend analysis of time series data, visualization of alliance structures based on network analysis, automatic learning of network attack patterns, and calculation of risk scores, it is possible to achieve high-precision and efficient franchise store review and management that was previously difficult to achieve manually. Specifically, this system containerizes each functional module on the server side as an independent microservice. It uses container orchestration tools to dynamically allocate computing resources based on the load of each microservice. An application programming interface (API) gateway routes all requests from the terminal to the appropriate microservice. To accelerate the inference processing of generative AI, a dedicated computing cluster equipped with multiple parallel processors is built. As a generative AI model, the system employs a large language model with billions to hundreds of billions of parameters and a multimodal model capable of processing both images and text simultaneously. As input data, the system acquires text data of the franchise store's business plan and a tensor of the store's exterior image. As a specific example of input data, the system accepts 5,000 words of text about future business development and RGB image data with a resolution of 124 pixels square. As output data, the system generates a score representing the likelihood of business success and extracts equipment information from the image as structured data. As a specific example of output data, the system outputs a likelihood score of 0.92 and a text label indicating the presence of security cameras. This system, as a subsequent processing step, sends the output data to the database management department and utilizes it to calculate the comprehensive evaluation score of franchise stores. In document parsing based on natural language processing, this system leverages the self-attention mechanism of the Transformer architecture to accurately capture the context of long texts. In image recognition, this system uses convolutional neural networks to extract subtle features such as imprints and watermark patterns on certificates. In anomaly detection, this system uses a latent spatial model learned from past normal transaction data to calculate the reconstruction error of new data and classifies data with errors exceeding a threshold as risky.This system connects multiple AI models as a pipeline, performing multi-faceted analysis of a single input data in parallel. It processes massive amounts of information instantly, exceeding the limits of human cognitive ability. This system dramatically reduces the time required for business review from weeks to minutes. It eliminates biases inherent in human judgment, guaranteeing objective and homogeneous review quality. Specifically, this system is applied to the onboarding process of franchised stores for globally operating settlement agency operators and to the buyer evaluation system for manufacturing companies with multiple suppliers. This system leverages its technological advantages to enhance the processing efficiency of computer technology itself, providing a high-level information processing foundation.

[0085] (System Composition) The system involved in this implementation includes an information acquisition department, a judgment department, a mapping department, a database management department, an external linkage department, an AI learning management department, a log management department, a security management department, a user interface department, an input support department, a result display department, a notification department, a file upload department, a download department, and a user authentication department. The information acquisition department receives from terminals the following information of potential franchisees: company name, representative name, location, date of establishment, capital, number of employees, business scope, financial statements for the past three periods, all historical certificates, regulations regarding personal information protection, pledges to exclude antisocial forces, past records of illegal transactions and administrative guidance, company website URL, industry classification, and technical information used for settlement system linkage (API specifications, security certification information, system architecture diagram, communication protocol specifications, encryption method descriptions, sandbox test results, etc.). It also automatically collects relevant information from the company's existing franchisee database, past review history, transaction performance database, legal department's contract database, accounting department's financial database, system linkage history database, and audit record database. For example, the Information Acquisition Department searches using company name, representative name, location, and establishment date as keys to verify information with existing franchisees. It extracts details from past review histories and transaction records, including records of irregularities and disputes, contract terminations and business cessations, and audit findings. Furthermore, through external collaboration, it automatically obtains credit ratings and financial scores from credit investigation agency databases, administrative guidance records and law violation records from publicly available information from government agencies, reputation and word-of-mouth from company information websites and news websites, the frequency of company name occurrences and co-occurrence with negative terms from social networking sites (SNS), membership and industry reputation from industry associations, intellectual property holding and application status from patent information databases, and past system incident information from cybersecurity vulnerability databases. For instance, it automatically collects information such as the risk of online criticism of company names on SNS, past administrative guidance records, expulsion history from industry associations, number of patent applications and patent content, history of cyberattacks, and the existence of vulnerability reports. Specifically, this information acquisition department installs a distributed crawler system that asynchronously and in parallel collects data from diverse data sources. This department passes the input data received from the terminal to a normalization module. This normalization module performs standardization of full-width and half-width characters, removes unnecessary whitespace, and standardizes date formats. Using the normalized company name and representative name as search keys, this department issues SQL queries to the company's internal database. This department uses Levinstein distance or N-gram as a fuzzy search algorithm to quickly identify similar companies with fluctuating transaction values. This department sends data acquisition requests to external collaboration departments. These external collaboration departments then call the RESTful application programming interface provided by the credit investigation agency.This external liaison department uses the OAuth protocol as the authentication method to establish a secure communication path. For government websites, this department controls a headless browser to parse dynamically generated document object model trees and extract necessary text data. This information acquisition department connects to a streaming application programming interface (API) to collect submission data in real time from social network services. This department applies natural language processing to the collected text data, using a morphological analysis engine to segment it into words. This department compares the data with a dictionary of words with negative sentiment polarity, counting the co-occurrence frequency with company names. This department parses patent gazette data provided in XML format from a patent information database, extracting applicant names, invention names, and international patent classification codes. This department retrieves JSON-formatted data containing common vulnerability identifiers from cybersecurity-related databases. This department stores all collected information in a graph database with companies as nodes and information relationships as edges. Through these technical processes, this system improves the comprehensiveness and speed of data collection by several orders of magnitude compared to manual searching and collection by humans. This system prevents information silos by constructing an integrated knowledge graph from scattered data both inside and outside the company. Specifically, it is applied in automated information collection systems within corporate credit management departments and due diligence support systems in investment funds. The system effectively eliminates bottlenecks in information collection during the initial stages of review, providing high-quality input data for subsequent AI-driven decision-making.

[0086] The Judgment Department inputs all information collected by the Information Acquisition Department into the Generative AI, performing various information analysis and judgments. For example, regarding financial information, the Generative AI automatically reads financial statements from the past three periods, automatically calculating sales, profit, equity ratio, whether there is insolvency, cash flow health, accounts receivable and payable turnover periods, fixed asset increases and decreases, investment activities, short-term and long-term borrowings, inventory increases and decreases, and whether there are late taxes, etc., and compares it with industry averages or past franchise data to score financial health. Regarding all historical documents, the Generative AI uses OCR or natural language processing to automatically parse the content, extracting information such as the legitimacy of legal entity status, history of representative changes, contents recorded in the purpose column, history of capital increases and decreases, history of location changes, records of dissolution or liquidation, mergers, divisions, and business transfers, and compares it with past violations or risk patterns. Regarding personal information protection systems, generative AI analyzes checklists or regulatory documents to automatically determine their suitability with laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, countermeasures against information leaks, the status of third-party certifications, whether there is entrusted management, and the completeness of personal information management ledgers. Regarding the exclusion of antisocial forces, the system assesses risk by comparing the content of pledges with external databases, such as comparing against antisocial force lists or verifying past administrative penalty records. Regarding past records of illegal transactions and administrative guidance, generative AI crawls external databases or news reports to extract past administrative penalties or dispute cases, litigation history, content of administrative guidance, whether there are improvement reports, the implementation status of countermeasures to prevent recurrence, the frequency or severity of administrative guidance, and assesses the risk level. Regarding the content recorded on corporate websites, generative AI automatically crawls the website to automatically determine the authenticity of the recorded content, whether there is exaggerated advertising, differences from past recorded content, the accuracy of product or service descriptions, whether there are any legal violations, whether there is any infringement of third-party rights, and whether there is any false representation or violation of the Gift Representation Act. Regarding industry risks, AI determines whether an industry is high-risk by referencing factors such as the overall industry violation rate or administrative guidance rate, past franchisee dispute rate, expulsion rate from industry associations, industry-specific regulatory risks, and social creditworthiness. Regarding the suitability of settlement system integration, AI analyzes technical information or API specifications to automatically determine compatibility with existing settlement systems, compliance with security requirements, encryption strength, communication protocol standard compliance, sandbox testing pass / fail status, past system failure history, and the presence of external vulnerability reports. For example, by automatically comparing the API specification with the settlement service provider's standards, AI automatically determines compatibility and security vulnerabilities, and analyzes sandbox test logs to detect any unauthorized communication or abnormal responses.Specific examples of prompts used by generative AI include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal status and past risk factors from the content of all historical certificates and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following franchise website and determine whether there are any false records or exaggerated advertisements," "Compare the following API specifications and settlement system requirements and determine the suitability of the linkage," and "Analyze the following sandbox test logs and detect any abnormal responses or unauthorized communications," etc. Specifically, this decision-making unit uses a large-scale language model based on the Transformer architecture. As input data, this decision-making unit obtains numerical data from financial statements or text data from all historical certificates. As preprocessing, this decision-making unit segments and lexicalizes the text data into sub-words. This decision-making unit converts the lexical sequence into multi-dimensional embedding vectors. For example, this decision-making unit generates 512-dimensional word embedding vectors. As a specific example of input data, this decision-making unit receives numerical data such as sales revenue of 100 million yen, or text sequences containing information representing change history. This decision-making unit feeds these input vectors into a self-attention mechanism. This decision-making unit calculates the dependencies between each word, obtaining a feature expression that takes into account the context. As output data, this decision-making unit generates a probability distribution representing the risk level, or a continuous value score representing financial soundness. As a specific example of output data, this decision-making unit outputs a value such as 0.15 representing the probability of bankruptcy, or a probability value such as 0.01 representing the association with antisocial forces. This decision-making unit further outputs a text word sequence in natural language explaining the basis for the decision. As a subsequent processing step, this decision-making unit determines whether the output probability value exceeds a pre-set threshold. If it exceeds the threshold, this decision-making unit executes a branch processing step of adding the franchise to the list requiring review. This decision-making unit sends the output score to the mapping unit. In the parsing of financial information, this decision-making unit represents the extracted numerical values ​​as multi-dimensional vectors and calculates the cosine similarity with the financial vectors of past bankrupt companies. In analyzing all historical certificates, this assessment department uses an optical character recognition (OCR) model to extract text from image tensors and a named entity recognition (NAME) model to extract company names, personal names, and dates as structured data. In analyzing personal information protection regulations, this assessment department calculates the semantic similarity between the text vectors of regulatory documents and the text vectors of legal guidelines, and identifies deficient clauses. In analyzing websites, this assessment department uses a multimodal model of images and text to detect inconsistencies between product images and descriptions. In assessing the suitability of checkout system integration, this assessment department inputs the text data of the instruction manual and the time-series data of the test logs into a recurrent neural network to detect sequences of abnormal communication patterns.This system utilizes advanced computational processing to parse massive documents with a speed and precision impossible for human visual verification. Its generalization capabilities detect ambiguous patterns or unknown risk models that rule-based programs cannot handle. Specifically, this system is applied to advanced compliance inspection systems in financial institutions or automated review engines for platform operators. This system provides core technological improvements for fully automated review processes.

[0087] The Mapping Department automatically categorizes franchisees based on various attributes, including the appropriate judgment and risk assessment results from the Judgment Department, the franchisee's industry, region, transaction scale, risk level, settlement system linkage status, presence of administrative guidance records, intellectual property ownership status, number of employees, capital size, years of establishment, cybersecurity risk level, and past system failure history. It then generates a franchisee map using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, timelines, risk heatmaps, and cybersecurity risk maps. The results are displayed on the terminal side, allowing department heads to refer to them in real time. For example, the risk management manager can use the heatmap to identify the distribution of franchisees in high-risk industries, the sales manager can use the map to identify the distribution of franchisees in newly expanded areas, the legal manager can use the timeline to track the timeline of franchisees with administrative guidance records, and the system manager can use the risk map to identify franchisees with high cybersecurity risks. Specifically, this Mapping Department executes a rendering pipeline that converts multidimensional data structures into visual elements. This Mapping Department receives various scores or attribute information output from the Judgment Department as structured data in JSON format. This mapping department unfolds the structured data in memory to build an in-memory database. In map visualization, it uses a geocoding API to convert the location information of franchisees into latitude and longitude coordinates. It utilizes a web graphics library to draw map tiles on the browser's canvas elements. Based on the coordinate data, it overlays markers or polygons on the map. In heatmap generation, it calculates the sum of the density or risk scores of franchisees existing within a specific geographic area. It maps the calculation results to two-dimensional grid data and uses kernel density estimation to generate smooth color gradients. In network diagram generation, it applies a layout algorithm based on a mechanics model. It sets attraction and repulsion forces between nodes and iteratively calculates the stable configuration with minimum energy. In tree diagram generation, it executes an algorithm that divides the hierarchical data into rectangular area ratios. In timeline generation, it defines the timeline as a one-dimensional coordinate system and configures plotting points according to the occurrence time of events. This mapping unit actively utilizes the hardware acceleration of the graphics processing unit on the terminal side during these rendering processes. Even with large datasets containing tens of thousands of data points, this mapping unit can achieve smooth panning or zooming operations without reducing the frame rate. This mapping unit processes filtering operations from the user in real time. For example, if the risk score threshold is changed via a slider control, this mapping unit immediately queries the in-memory database to exclude data points that do not meet the criteria from the rendering objects. Through these techniques, this system transforms complex multidimensional data into a form easily processed by the human visual cortex. This system enables the intuitive discovery of patterns or outliers hidden within the data.This system eliminates the cognitive load of reading large amounts of text data, resulting in faster decision-making. Specific applications include dashboard systems for management, regional analytics tools for marketing departments, and threat visualization systems for security operations centers. This system facilitates information sharing between departments and supports data-driven organizational operations.

[0088] The Database Management Department manages all franchisee information, judgment results, scoring data, mapping data, AI learning data, user operation logs, external linkage history, and cybersecurity incident information in a unified manner, and saves this information as an audit trail. For example, it saves information on all previously registered franchisees, scoring results output by the Judgment Department, franchisee maps generated by the Mapping Department, learning data accumulated by the AI ​​Learning Management Department, all user operation history, external data acquisition history obtained by the External Linkage Department, and cybersecurity incident occurrence history in chronological order, enabling searching, retrieval, and reuse as needed. Specifically, the Database Management Department employs a distributed database architecture capable of handling petabyte-scale data volumes. It implements sharding technology, dividing and configuring data across multiple physical servers. The department inputs franchisee identifiers into a hash function, determining the shards for storing data based on the output value. This prevents concentrated access to specific servers, allowing read / write performance to scale horizontally. To improve data availability, the department asynchronously replicates data from each shard to multiple replica nodes. This database management unit possesses a failover mechanism that automatically promotes auxiliary nodes to become the primary node in the event of a primary node failure. It utilizes a time-series database engine specifically designed for storing time-series data. This unit writes timestamped data, such as user operation logs, system linkage history, and network security incident history, as continuous blocks along a timeline to disk. This enables extremely high-speed processing of queries specifying range searches for specific periods or trend analysis over time. To optimize storage costs, this unit automates data lifecycle management. It configures frequently accessed, recent data on high-speed solid-state drives. It performs tiered storage management, automatically moving past audit trail data over a certain period to inexpensive, high-capacity hard drives or cloud-based archive storage. To ensure data integrity, it periodically takes snapshots of the entire database and uses checksums to detect data corruption. In its search and retrieval functions, this unit supports complex join and aggregation processing based on Structured Query Language. It uses a query optimizer to dynamically optimize execution plans and select the access path that returns results in the shortest time. This system securely and efficiently accumulates vast information assets through these technologies. It eliminates data silos, providing a foundation for company-wide data utilization. This system effectively minimizes downtime and ensures business continuity. Specifically, this system is applied to comprehensive customer databases in financial institutions or transaction history management systems for large-scale platforms. While meeting stringent audit requirements, this system provides a robust foundation for big data analytics.

[0089] The External Liaison Department collaborates with multiple external information sources, including credit investigation agencies, government agencies, corporate websites, news websites, social networking sites (SNS), industry associations, patent databases, and cybersecurity vulnerability databases. Through various methods such as APIs, web scraping, RSS subscriptions, and database integration, it automatically acquires necessary information. For example, it obtains credit ratings or financial scores from credit investigation agencies via API integration; administrative guidance records or law violation records from government agencies via web scraping; the frequency of company names or co-occurrence rates with negative terms from SNS; membership qualifications or industry reputation from industry associations; intellectual property ownership or application status from patent databases; and past system event information or vulnerability reports from cybersecurity vulnerability databases. Specifically, this External Liaison Department is equipped with a comprehensive interface module that handles various protocols and data formats. In application programming interface (API) integration, it adds an authentication token to the HTTP request header to establish a secure communication channel. As a response from an external server, it receives data in JSON or XML format. The department then parses the received data into its internal standardized data structure. In web crawling, this external coordination unit automatically parses the target website's robot exclusion protocol, allowing crawler operations within permitted limits. To reduce the load on the target server, this external coordination unit implements a rate-limiting avoidance algorithm that inserts random delays between requests. For dynamically generated single-page applications, this external coordination unit launches a headless browser and extracts the Document Object Model (DOM) tree while waiting for JavaScript execution to complete. From the extracted HTML data, this external coordination unit uses XPath or CSS selectors to accurately extract specific text nodes. In RSS subscription coordination, this external coordination unit periodically polls the subscribed URLs, comparing update dates and times to obtain new reports only by difference. This external coordination unit performs data cleaning on the acquired unstructured text data. This external coordination unit uses regular expressions to remove HTML tags or special characters. This external coordination unit unifies character encoding to UTF-8. In data acquisition from social networking services, this external coordination unit maintains streaming connections, filters submissions containing specific keywords in real time, and accumulates them in a memory buffer. When the buffer reaches a certain amount, this external coordination unit writes the data to the database as a batch. In the event of communication errors or timeouts, this external linkage unit employs an exponential backoff algorithm, gradually increasing the retry interval while performing retry processing. Through these techniques, the system integrates numerous external data sources with varying formats and acquisition methods into a single system. This system transcends the limitations of manual information gathering, dramatically improving the freshness and comprehensiveness of information. Furthermore, this system delivers the technical benefit of a loosely coupled architecture that allows for flexible adaptation to changes in external system specifications.This system is specifically applied in enterprise open-source intelligence gathering systems or as an automated competitive analysis tool for marketing departments. It monitors changes in the external environment in real time, supporting rapid decision-making.

[0090] The AI ​​Learning Management Department continuously learns from past judgments made by the Judgment Department, new information acquired from external sources, the occurrence of disputes after actual transactions, administrative guidance, franchisee bankruptcies or mergers, growth performance, cyberattacks, and system failures to improve the accuracy of generative AI judgments. For example, if a franchisee previously deemed unsuitable later causes problems, this case is used as learning data and reflected in future judgment benchmarks. The AI ​​Learning Management Department also automatically acquires information and reflects it in the Judgment Department's algorithm in the event of new legal amendments or changes in industry guidelines. The AI ​​Learning Management Department compares past judgment results with actual post-event risk occurrences and continuously optimizes the judgment logic. For example, it learns from past system linkage information or vulnerability reports of franchisees that have experienced cyberattacks and reflects this in future cyber risk judgments. Specifically, this AI Learning Management Department constructs a pipeline that automates the application of machine learning. This AI Learning Management Department regularly extracts past judgment logs and post-event results data from the Database Management Department. The AI ​​Learning Management Department inputs the extracted data into the Feature Engineering module. This module automatically performs missing value completion, outlier removal, categorical variable encoding, and continuous variable scaling. The AI ​​Learning Management Department then splits the preprocessed data into training, validation, and test datasets. As input data, the AI ​​Learning Management Department obtains attribute vectors or text embedding vectors for franchise stores. Specific examples of input data include numerical values ​​such as the number of employees (e.g., fifty) or text vectors representing past system failure histories. As output data, the AI ​​Learning Management Department defines continuous values ​​representing the probability of risk occurrence or categorical labels. Specific examples of output data include numerical values ​​such as the probability of system failure occurrence (e.g., 0.75) or labels such as "high risk." The AI ​​Learning Management Department uses a deep learning framework to perform relearning of the neural network. It uses gradient descent to update weight parameters in a way that minimizes the loss function. In hyperparameter optimization, the AI ​​Learning Management Department uses a Bayesian optimization algorithm to efficiently search the search space and determine the optimal learning rate or batch size. Upon receiving new legal amendment information, our AI Learning Management Department updates the vector index of the external knowledge base using retrieval-enhanced generation techniques. This allows us to dynamically assemble the latest knowledge into prompts during inference without altering the model's weights. As a subsequent processing step, we evaluate the performance of the relearned model using a test dataset. We calculate evaluation metrics such as accuracy, recall, and F-score. We deploy the new model to the production inference endpoint only if its performance surpasses that of the old model. During deployment, we employ a blue-green deployment approach to minimize system downtime.This system employs these technologies to autonomously adapt to environmental changes or new threats. It effectively prevents model accuracy degradation and maintains high predictive performance over the long term. Specifically, this system is applied to violation detection systems in financial institutions or as the threat intelligence foundation for cybersecurity products. While minimizing human intervention, the system provides advanced decision-making logic that consistently reflects the latest insights.

[0091] The Log Management Department meticulously records all operation history, data acquisition history, judgment history, mapping history, system linkage history, and network security incident history, enabling auditing and compliance departments to refer to them later. For example, it records in chronological order who input which franchisee's information, when, what judgment result was output, when which external information was acquired, what mapping was generated, and what network security incident occurred, saving this as an audit trail. Specifically, this Log Management Department installs a comprehensive log collection architecture in a distributed system environment. Lightweight log collection agents reside on each functional module or server node. These agents monitor the application's output log files in real time, reading them whenever new log entries are appended. The department converts the read log data into a structured JSON format. It assigns metadata such as timestamps, event IDs, importance levels, source IP addresses, user identifiers, and the names of executed functions to the log data. The department then sends the structured log data to a message broker. This message broker functions as a high-throughput distributed messaging system, temporarily buffering large amounts of log data. This log management unit asynchronously consumes log data from the message broker, indexes it, and stores it in the full-text search engine cluster. Within the full-text search engine, this unit builds an inverted index, enabling ultra-fast searches based on arbitrary keywords or conditions. To completely prevent log tampering, this unit applies blockchain or hash chain technology. At regular intervals, this unit generates blocks of log data and calculates the cryptographic hash value of each block. By including the calculated hash value in the header of the next block, this unit records past log data as an immutable chain structure. This unit provides a dedicated interface for the auditing department. Through this interface, this unit supports the execution of complex search queries or the extraction of specific event patterns. For example, this unit automatically detects events such as specific users downloading large amounts of data outside of business hours through correlation analysis and displays them as alerts. Through these technologies, the system maximizes its transparency and traceability. This system provides a specific forensic basis for rapidly identifying the cause and scope of damage in the event of a security incident. This system technically guarantees compliance requirements, resulting in a reduction of legal risks for enterprises. Specifically, it is applied to audit response systems based on financial instrument transaction laws, or to electronic medical record access log management systems in medical institutions. This system instantly extracts meaningful information from massive amounts of log data, dramatically improving the efficiency of audit operations.

[0092] The Security Management Department implements data encryption, access control management, unauthorized access detection, data backup, fault recovery functions, network attack detection, and abnormal communication detection to ensure the security and reliability of the entire system. For example, access permissions are meticulously set for each user, and information that can be viewed and edited is controlled by each department. All communications are encrypted, alarms are automatically issued upon detecting unauthorized access, and data backups are regularly performed, enabling rapid recovery in the event of a failure. Specifically, the Security Management Department implements a comprehensive security framework based on a multi-layered defense architecture. In data encryption, advanced encryption standards such as AES-256 are applied to stored data, performing encryption at the block or file level. For communication data, the Security Management Department uses the TLS 1.3 protocol, employing forward-secrecy cipher suites to establish end-to-end encrypted communication. In access control management, a role-based access control model is adopted. The Security Management Department dynamically issues permission tokens based on the user's position or department, and performs token signature verification when requesting access to various microservices. This security management department integrates network-based and host-based intrusion detection systems for unauthorized access and network attack detection. It employs machine learning-based behavioral detection algorithms. As input data, it acquires time-series tensors such as packet header information, payload statistical characteristics, and API call frequency from network traffic. It uses autoencoders or recurrent neural networks to learn models of normal communication patterns. The department calculates the reconstruction error between real-time input data and model predictions, detecting data with errors exceeding a threshold as zero-day attacks or anomalous communication. As follow-up to detected anomalies, it dynamically updates firewall rules to block communication from the affected IP address and sends an emergency alert to the administrator. In data backup, it automatically executes plans combining differential and full backups. Backup data is asynchronously transferred to multiple geographically dispersed data centers for disaster recovery. In failure recovery, it utilizes the self-healing capabilities of container orchestration tools. It automatically discards container instances that fail health checks and maintains service availability by restarting new instances. This system minimizes the risk of information leakage caused by external cyberattacks or internal crimes through these technical processes. It effectively eliminates single points of failure and achieves high fault tolerance. Specifically, this system is applied to the core systems of enterprises undertaking critical infrastructure or the backend systems of cloud services storing vast amounts of personal information. Based on the principle of zero trust, this system consistently verifies all access, providing a robust and secure environment.

[0093] The user interface department operates on the terminal side, comprehensively providing input support or auto-completion functions, file upload functions, result display functions, notification functions, download functions, user authentication functions, etc. For example, when a company name is entered, it automatically displays a list of past candidate names or similar company names; when uploading financial statements, it automatically performs format conversion; it displays easily understandable information such as suitable or unsuitable judgment results sent from the judgment department, visualization information of franchise store maps, AI-based summary comments, risk alerts, requests for additional information, and network security risk alerts, and allows users to download them in PDF or Excel format as needed. The notification department notifies users in real time of judgment results or requests for additional information, risk alerts, feedback on AI learning results, and emergency notifications in the event of network incidents. The file upload department automatically determines the format of uploaded files, extracts text from PDFs, and performs OCR processing on images as needed to standardize the data. The download department enables users to download judgment results, franchise store maps, AI-based summary comments, etc., in various file formats. The user authentication department has security functions such as single sign-on or multi-factor authentication, and controls access permissions by department. Specifically, this user interface unit is built as a high-level front-end application operating on the rendering engine of a web browser. It employs a virtual document object model, optimizing rendering performance by minimizing differential updates to screen redraws corresponding to state changes. The user interface unit's input support monitors keystroke events, applies debouncing, and sends asynchronous requests to the server when input stops for a certain period. This prevents excessive load on the server while achieving smooth auto-completion. The file upload unit utilizes the HTML5 file API to read the binary data of the user-selected file into memory. It parses the file header information to accurately determine the MIME type. For large files, it divides the data into multi-megabyte chunks and sends them in parallel as multiple form data parts. It leverages the terminal's computing power to perform preprocessing, reducing the resolution of image files before sending, saving communication bandwidth. The results display unit parses the JSON-formatted judgment result data received from the server and dynamically assigns values ​​to various components on the screen using data binding technology. Upon receiving the visualization information of the franchise store map, the results display unit invokes a web graphics library to perform hardware-accelerated rendering on the canvas elements. The notification unit maintains a constant WebSocket connection with the server, establishing a full-duplex communication channel. The notification unit immediately receives event messages pushed from the server and invokes the operating system's native notification API to display desktop notifications. The download unit converts the binary data received from the server into Blob objects, dynamically generates the download attributes of anchor tags, and triggers a file save dialog box.This user authentication department uses the OpenID Connect protocol to securely exchange JSON WebTokens with external authentication providers. Through these technologies, the system offers operability and responsiveness comparable to desktop applications. It hides the effects of network latency, maximizing user productivity. Specifically, this system is applicable to back-end business systems with complex input data and continuous verification processes, or monitoring dashboards requiring real-time status updates. It efficiently utilizes terminal-side resources to achieve a distributed processing architecture that distributes server-side load.

[0094] In this way, the various parts of the system involved in this implementation work in tandem to achieve high-precision and efficient franchise store review and management centered on generative AI. Specific examples of prompts that the generative AI can read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal entity status and past risk factors from the contents of all the following historical certificates and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following franchise store websites and determine whether there are any false records or exaggerated advertisements," "Compare the following API specifications and settlement system requirements and determine the suitability of the linkage," "Analyze the following sandbox test logs and detect any abnormal responses or unauthorized communications," etc. By using these prompts, the generative AI performs various functions such as information analysis, judgment, summarization, comparison, visualization, and network risk assessment, achieving high-precision and efficient franchise store review and management that was previously difficult to achieve manually. Specifically, this system installs a prompt engineering module that dynamically generates and optimizes these prompts. This prompt word engineering module converts data in various formats collected by the information acquisition department into text sequences that can be processed by large-scale language models. To ensure that the number of lexical units in the input data does not exceed the upper limit of the model's context window, this module performs automatic summarization or discarding of low-importance information in chunks. In the generative AI inference pipeline, this system integrates multiple specialized models. For example, in the parsing of financial data, this system assigns a model specifically for numerical calculation and logical reasoning; in the parsing of specified documents, it assigns a model specifically for contextual understanding of long texts. As input data, this system acquires a combined text vector of prompt words and target data. As a specific example of input data, this system receives a combination of numerical arrays of items from financial statements and text indicating scoring. As output data, this system generates response text or probability distribution tensors for each task. As a specific example of output data, this system outputs a numerical value such as a financial soundness score of 0.88 or a classification label such as "abnormal communication detected." As subsequent processing, this system aggregates the outputs from each model into the integration module to calculate the final comprehensive judgment score. To minimize inference processing latency, this system employs quantization techniques to convert model weight parameters from floating-point numbers to low-bit integers. It leverages the tensor cores of parallel processing processors to perform matrix multiplication operations at high speed. These techniques do not simply mimic human cognitive processes; rather, they utilize mathematical operations in high-dimensional space to process information at a scale and speed impossible for humans. This system effectively eliminates the human element in review processes and establishes highly reproducible and objective evaluation benchmarks.This system is specifically applied in the real-time franchisee onboarding system of fintech companies or the supply chain risk management system of global enterprises. It provides the foundational technology for building a robust risk management system without hindering business growth.

[0095] (Implementation steps) Step 1: Input and upload of candidate franchise store information. Using the user interface unit on the terminal side, input or upload the candidate franchise store's enterprise name, representative name, location, date of establishment, capital amount, number of employees, business content, financial statements for the past 3 periods (balance sheet, income statement, cash flow statement), full certificate of historical matters (PDF file or image file), regulatory documents on the personal information protection system (privacy policy, information management regulations, employee education records), pledge letter for excluding anti-social forces, past records of illegal transactions and administrative guidance, URL of the enterprise website, industry classification, and technical information for settlement system linkage (API specification, security authentication information, system structure diagram, communication protocol specifications, description of encryption methods, sandbox test results, etc.). The input support unit performs auto-completion of the enterprise name and representative name, postal code conversion for the location, format conversion of financial statements, OCR processing of the full certificate of historical matters, checklist conversion of the personal information protection system, automatic analysis of the API specification, etc., to improve the efficiency of input operations. Specifically, in this step, the user interface unit on the terminal side performs real-time verification processing on the input text data. This user interface unit uses regular expressions to verify the validity of the formats of email addresses and telephone numbers. This user interface unit confirms through purification processing whether invalid character encodings or script tags are included. This file upload unit reads the files dragged and dropped by the user into the memory. This file upload unit calculates the hash value of the file, and determines on the client side whether it is a duplicate of an already uploaded file. When an image file is uploaded, this file upload unit uses the computing resources on the terminal side to perform image preprocessing such as edge extraction and contrast adjustment. This file upload unit sends the preprocessed image data to the information acquisition unit on the server side through asynchronous communication. The information acquisition unit on the server side inputs the received image data into the optical character recognition model. This optical character recognition model has an architecture combining a convolutional neural network and a recurrent neural network. This optical character recognition model acquires, as input data, an image tensor composed of a two-dimensional arrangement of pixel values. As a specific example of input data, this optical character recognition model receives a grayscale image tensor with a resolution of 2048 pixels square. This optical character recognition model generates, as output data, a sequence of recognized characters, the bounding box coordinates of each character, and a confidence score. As a specific example of output data, this optical character recognition model outputs the character string "Kabushiki Gaisha" and a numerical value of 0.99 as the confidence level. As subsequent processing, this information acquisition unit identifies character regions whose confidence scores are lower than a threshold, and returns a warning message to the user interface unit to prompt manual confirmation. To extract the table structure from the PDF file of the financial statements, this information acquisition unit uses a heuristic line drawing detection algorithm and a machine learning-based layout analysis model. This information acquisition unit converts the extracted numerical values into a structured JSON format and stores them in the database management unit.This system automates the process of converting unstructured data into machine-readable structured data through these technologies. It significantly reduces the human cost and time required for data input and eliminates input errors. Specifically, this system is suitable for back-office operations in financial institutions handling large volumes of paper or PDF documents, as well as electronic application processing systems in administrative agencies. This system lays the foundation for providing high-quality, low-noise input data for subsequent AI parsing pipelines.

[0096] Step 2: Automatically acquire information from internal and external databases and external information sources. In addition to information received from the terminal, the Information Acquisition Department automatically collects relevant information from the company's existing franchise database, past review history, transaction performance database, legal department's contract database, accounting and finance department's financial database, system linkage history database, and audit record database. Furthermore, the External Liaison Department automatically acquires credit ratings, administrative guidance records, reputation, word-of-mouth, news reports, online criticism history, membership status, intellectual property ownership, cyberattack victim records, and vulnerability reports from databases of credit investigation agencies, publicly available information from administrative agencies, corporate information websites, news websites, social networking sites, industry association lists, patent information databases, and cybersecurity-related vulnerability information databases. Specifically, in this step, the Information Acquisition Department and the External Liaison Department utilize a distributed processing framework to execute an advanced parallel data collection pipeline. The Information Acquisition Department uses the company name and representative name received from the terminal as seed information to dynamically generate queries targeting multiple data sources. When accessing internal databases, the Information Acquisition Department uses connection pooling technology to minimize database connection overhead. This information acquisition department asynchronously issues complex combined queries across multiple tables, expanding the result set into data frames on memory. This external linkage department sends non-blocking HTTP requests using asynchronous I / O in parallel to external application programming interfaces (APIs). This external linkage department monitors rate limits for each external source and dynamically controls the request sending frequency using a token bucket algorithm. In web crawling, this external linkage department orchestrates a distributed group of crawler nodes, allocating crawling tasks according to the domain name of the target website. This external linkage department uses natural language processing to extract named entities from the acquired HTML documents, excluding noisy advertising and navigation elements. This information acquisition department performs entity parsing processing on the massive amounts of data collected from inside and outside the company. This information acquisition department considers factors such as differences in company names and abbreviations, and changes in location due to relocation, to determine whether records obtained from different data sources point to the same company. In this determination, this information acquisition department calculates the Levenstein distance and Jaro-Winkler distance between the strings of each attribute, and uses a machine learning-based classifier to calculate the identity score. This information acquisition department merges groups of records with identity scores exceeding a threshold into a single master record. This information acquisition unit inputs the collected text data into a vectorization model, converting it into high-dimensional embedding vectors. These vectors are then stored in a vector database, accelerating subsequent similarity searches. Through these techniques, the system breaks down information silos, reconstructing scattered, heterogeneous data into a comprehensive knowledge base. This system completely eliminates the need for manual information gathering using browsers or search engines. This system delivers a significant technological boost to the comprehensiveness and speed of information collection.This system is applicable to automated investigation systems in corporate credit management departments and data aggregation foundations used in investment banks for corporate valuation. It eliminates information asymmetry in the initial stages of review, enabling more accurate risk assessment.

[0097] Step 3: Multi-dimensional information analysis and judgment through generative AI. The judgment department inputs all information collected by the information acquisition department into the generative AI for analysis and judgment of various information. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profit, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, the increase or decrease of fixed assets, the content of investment activities, the shift of short-term and long-term loans, the increase or decrease of inventory, and whether there are any tax arrears, etc., and compares them with industry averages or past franchise store data to score the soundness of finances. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extract the legitimacy of legal person status, the history of changes in representatives, the content recorded in the purpose column, the history of increases or decreases in capital, the history of changes in location, records of dissolution or liquidation, mergers, divisions, business transfers, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection mechanisms, generative AI analyzes checklists or regulatory documents to automatically determine their suitability to laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, the existence of countermeasures against information leaks, the status of third-party certifications, whether there is entrusted management, and the completeness of personal information management ledgers. Regarding the exclusion of antisocial forces, the presence of risk is determined by comparing the content of pledges or external databases, such as comparing with antisocial force lists or verifying past administrative penalty histories. Regarding past records of illegal transactions and administrative guidance, generative AI retrieves data from external databases or news reports, extracting past administrative penalties or dispute cases, litigation history, the content of administrative guidance, whether there are improvement reports, the implementation status of countermeasures to prevent recurrence, the frequency or severity of administrative guidance, etc., to assess the risk level. Regarding the content recorded on the company website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, discrepancies with past recorded content, the accuracy of product and service descriptions, whether there are any violations of laws and regulations, whether there are any infringements of third-party rights, and whether there are any false representations or violations of the Goods Representation Act. Regarding industry risks, AI determines whether an industry is high-risk by referring to the overall industry violation rate or administrative guidance rate, past franchise dispute rate, expulsion rate from industry groups, industry-specific regulatory risks, and social credit rating. Regarding the suitability of the settlement system integration, technical information or API specifications are analyzed to automatically determine the compatibility or security requirements of the existing settlement system, the strength of encryption methods, the standard compliance of communication protocols, whether the sandbox test is passed, past system failure history, and whether there are any external vulnerability reports.Specific examples of prompts for generative AI to read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal entity status and past risk factors from the contents of all certificates for the following historical events, and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following franchise store websites and determine whether there are any false records or exaggerated advertisements," "Compare the following API specifications and settlement system requirements and determine the suitability of the linkage," and "Analyze the following sandbox test logs and detect any abnormal responses or unauthorized communications." Specifically, in this step, this decision unit executes an inference pipeline centered on a multimodal, large-scale language model. As input data, this decision unit simultaneously acquires a structured arrangement of financial values, text marker columns of regulatory documents, and a screenshot image tensor from the website. Specific examples of input data received by this decision unit include a arrangement of floating-point numbers representing the three-year trend of sales, the full text of a privacy policy, and an RGB image tensor with a resolution of 124 pixels square. This decision-making unit inputs these heterogeneous data into their respective encoders, transforming them into high-dimensional vector representations in a common latent space. This decision-making unit uses a cross-attention mechanism to calculate the correlation between text vectors and image vectors, fusing multimodal feature quantities. This decision-making unit inputs the fused feature vectors into the decoder network. As output data, this decision-making unit generates continuous values ​​of risk scores, classification labels, and natural language text tags explaining the basis for the judgment for each assessment item. As a specific example of the output data, this decision-making unit outputs a financial soundness score of 0.85, a classification label of low antisocial force risk, and a description of no unnatural increases or decreases in capital. As subsequent processing, this decision-making unit inputs the output scores into a weighted linear summation algorithm to calculate the final comprehensive judgment score. If the comprehensive judgment score is lower than a pre-set threshold, this decision-making unit performs branch processing to mark the franchise as unsuitable. During the inference process, this decision-making unit performs similarity searches on the vector index of an external knowledge base, dynamically obtaining relevant text vectors of past cases or the latest legal guidelines and appending them to the context of the prompts. This system utilizes these technologies to execute complex logical reasoning—which involves human experts verifying multiple data sources—as high-speed and accurate mathematical operations within the computer. It delivers the technical advantage of highly accurate extraction of hidden risk factors from unstructured data or ambiguous expressions that rule-based systems struggle with. Specifically, this system is applicable to advanced credit review systems for financial institutions or merchant screening systems for platform operators. It significantly shortens the lead time for review processes and builds a scalable risk management system.

[0098] Step 4: Generation and Notification of Judgment Results, Scoring, and Supporting Information. The judgment department scores each item, performs a comprehensive suitability judgment, and generates detailed scoring results or supporting data, AI-generated summary comments, comparisons with similar past cases, etc., which are then sent to the results display unit on the terminal side. The notification department notifies the user in real time of the judgment results or reasons for unsuitability, risk alerts, requests for additional information, feedback on AI learning results, and emergency notifications in the event of network incidents. For example, if financial soundness falls below a certain standard, there is a significant past record of administrative guidance, false information is found on the website, the personal information protection system is inadequate, antisocial forces are detected, the industry risk is high, there are technical problems with the settlement system linkage, the test fails in the sandbox test, or there are external vulnerability reports, the item is deemed unsuitable, and the reasons are automatically generated and notified to the terminal side. Specifically, in this step, the judgment department inputs the multi-dimensional score vector output by the generative AI into the aggregation algorithm. The judgment department multiplies the weight coefficient matrix corresponding to the importance of each evaluation item by the score vector to calculate the comprehensive suitability judgment score. In calculating this score, the decision-making unit employs a hybrid approach that integrates rule-based decision logic using a decision tree model with the output of generative AI. This unit combines the calculated overall score with the individual scores for each item, along with a natural language summary commentary from the generative AI output, to generate a structured result payload in JSON format. To generate comparisons with similar past cases, the unit calculates the cosine similarity between the current franchisee's feature vector and past franchisee vectors in the database. The unit extracts the highest-similar case and appends its result to the result payload. The unit publishes the generated result payload to a message queue system. The notification unit asynchronously subscribes to the result payload from the message queue. The notification unit parses the flags within the payload and assesses whether it contains unsuitable judgments or significant risk alerts. If the corresponding flags are present, the notification unit immediately sends a push notification data packet via a WebSocket connection established with the user interface unit on the terminal side. In highly urgent situations such as network events, in addition to the usual notification channels, the notification unit also invokes email sending APIs or SMS service APIs to perform follow-up processing, directly sending alerts to the administrator's mobile terminal. This system utilizes these technologies to structure complex decision-making results into a human-readable format and delivers them without delay. It enables real-time decision-making and rapid action, eliminating the need to wait for daily confirmation through batch processing. Specifically, this system is suitable for online transaction violation detection systems requiring second-level judgments, or mission-critical infrastructure monitoring systems.This system provides a robust notification mechanism directly related to the early detection of risks and the minimization of harm to victims.

[0099] Step 5: Automatic Generation and Visualization of Franchise Store Maps. The mapping department automatically classifies franchise stores based on multiple attributes, including the appropriate judgment and risk assessment results from the judgment department, the franchise store's industry, region, transaction scale, risk level, settlement system linkage status, presence of administrative guidance records, intellectual property ownership status, number of employees, capital scale, number of years established, network security risk level, and past system failure history. It generates franchise store maps using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, timelines, risk heatmaps, and network security risk maps. The results are displayed on the terminal side, allowing department heads to refer to them in real time. Specifically, in this step, the mapping department receives multi-dimensional attribute data and score data output from the judgment department as input. The mapping department expands this data as a multi-dimensional arrangement in memory. The mapping department executes clustering algorithms to automatically group franchise stores with similar attributes. For example, the mapping department uses the k-means algorithm to divide data points in high-dimensional space into multiple clusters. The mapping department calculates the centroid coordinates of each cluster and the distance between clusters. This mapping unit generates rendering data for visualization. When generating a network graph, it constructs a graph structure with franchisees as nodes and transaction relationships or similarities as edges. It performs layout calculations based on a mechanical model to determine two-dimensional coordinates that balance the attractive and repulsive forces between nodes. When generating a heatmap, it calculates a density function on a two-dimensional grid based on geographic coordinate data and generates a smooth, color-gradient pixel arrangement through kernel density estimation. The generated rendering data is serialized into JSON format and sent to the results display unit on the terminal. Based on the received data, the results display unit calls data-driven document manipulation libraries or web graphics libraries. Utilizing the terminal's graphics processing unit, it rapidly draws vector graphics or pixel data on canvas elements. It monitors mouse hover or click events from the user and performs interactive processing by displaying information about the data points corresponding to the coordinates of the events. Through these techniques, the system converts tens of thousands of complex, multi-dimensional data points into a format that the human visual cortex can instantly recognize patterns in. This system provides the intuitive ability to identify macro trends or micro anomalies that cannot be found through spreadsheet software. Specifically, it can be used as a portfolio analysis tool in business planning departments or a company-wide threat monitoring dashboard in risk management departments. This system strongly supports data-driven strategic decision-making and provides a foundation for improving data literacy across the entire organization.

[0100] Step 6: Continuously improve judgment accuracy through AI learning. The AI ​​Learning Management Department continuously learns from past judgments made by the Judgment Department, newly acquired external information, the occurrence of disputes after actual transactions, the occurrence of administrative guidance, the bankruptcy or merger of franchise stores, growth performance, the occurrence of cyberattacks, and the occurrence of system failures, thereby improving the judgment accuracy of generative AI. For example, if a franchise store previously deemed unsuitable later causes problems, this case is imported as learning data and reflected in future judgment benchmarks. In the event of new legal amendments or changes in industry guidelines, the AI ​​Learning Management Department also automatically acquires information and reflects it in the Judgment Department's algorithm. Specifically, this AI Learning Management Department constructs a continuous learning pipeline. This pipeline converts newly acquired data labeled with correct answers into multidimensional tensors. This AI Learning Management Department uses self-supervised learning and reinforcement learning to fine-tune the weight parameters of a large-scale language model. The system receives past judgment score vectors, time-series financial change matrices, and binary data indicating the occurrence of disputes as input data. This system projects these input data into a 512-dimensional embedding vector space. The AI ​​learning management unit minimizes the cross-entropy loss using backpropagation. The processing unit updates the weights of each layer of the neural network using gradient descent. As output, the system generates an updated attention weight matrix and a new risk assessment threshold set. This output is then used as initial parameters in the next inference process. The system inputs textual data of legal amendments into a natural language processing model. The model extracts the changed regulatory requirements from the text and outputs them as a rule-based constraint vector. This constraint vector is then assembled as a penalty term for the loss function. Thus, the system does not rely on manual updates by humans, but immediately reflects the latest legal requirements into the decision logic. This structure prevents model obsolescence. This structure leads to a continuous improvement in decision accuracy. This technology can be applied to fields requiring dynamic risk assessment, such as credit review systems in financial institutions or seller monitoring systems in e-commerce platforms.

[0101] (specific use case) For example, when an electronic settlement service operator, acting as a new franchisee, reviews a legal entity operating in the catering industry, the user interface of the terminal is used to input or upload the following information: company name, representative name, location, date of establishment, capital, number of employees, business scope, financial statements for the past three periods (balance sheet, profit and loss statement, cash flow statement), all historical documents (PDF file), regulations regarding the personal information protection system (privacy policy, information management regulations, employee training records), pledge to exclude antisocial forces, past records of illegal transactions and administrative guidance, company website URL, industry classification, and technical information used for the settlement system linkage (API specifications, security certification information, system architecture diagram, communication protocol specifications, encryption method descriptions, sandbox test results, etc.). The input support department automatically suggests past registered company names or similar names when inputting the company name, automatically completes the address based on the postal code when inputting the location, automatically extracts numerical data from PDF or Excel when uploading financial statements, performs OCR processing on the images of all historical documents, automatically generates a checklist for the personal information protection system, and also performs format conversion of the API specifications or automatic parsing of communication protocols. Specifically, this input support unit executes a preprocessing pipeline to extract structured data from unstructured data. This input support unit acquires uploaded PDF files or image data as two-dimensional pixel value tensors. This input support unit inputs this pixel value tensor into an optical character recognition model using a convolutional neural network. This model identifies character regions within the image as bounding boxes. This model outputs the characters within the identified regions as a sequence of character codes. This input support unit inputs the extracted text data into a natural language processing model. This model executes a named entity extraction algorithm. This model outputs attribute information such as company name, location, and capital as key-value structured data. This input support unit parses tabular data from financial statements. This input support unit extracts pairs of accounting items and amounts based on line information and spatial configuration of the text. This input support unit converts the extracted values ​​into a floating-point multidimensional permutation. This input support unit converts technical documents such as API specifications into syntax trees. This input support unit generates feature vectors representing the type of communication protocol or the strength of encryption methods. This system sends this structured data and feature vectors to the subsequent decision unit. This architecture eliminates manual input or visual verification by humans. It results in faster data entry and reduced input errors. This technology can be applied to the onboarding process of franchisees in online payment agency services, or to tenant registration systems for cloud services.

[0102] In addition to information received from terminals, the Information Acquisition Department automatically collects relevant information from various sources within the company, including existing franchisee databases, past review histories, transaction performance databases, legal department contract databases, accounting department financial databases, system linkage history databases, and audit record databases. For example, to verify information with existing franchisees, searches are performed using the company name or representative name, location, and establishment date as keys. From past review histories or transaction performance, searches are used to extract records of irregular transactions or disputes, contract terminations or transaction halts, and audit findings. Furthermore, through external collaboration, the department automatically obtains credit ratings or financial scores from credit investigation agency databases, administrative guidance records or legal violation records from publicly available information from administrative agencies, reputation or word-of-mouth from company information websites or news websites on the internet, the frequency of company name occurrences or co-occurrence rates with negative terms from social networking sites, membership qualifications or industry reputation from industry associations, intellectual property possession or application status from patent information databases, and past system incident information from cybersecurity vulnerability information databases. For example, the risk of online criticism of company names on social networking sites (SNS), past administrative guidance records, expulsion history from industry groups, number of patent applications or content of held patents, records of cyberattacks, or the existence of vulnerability reports are also automatically collected. Specifically, this information acquisition department drives a distributed data collection engine. This information acquisition department issues queries in parallel to multiple heterogeneous databases. This information acquisition department normalizes relational data obtained from internal company databases. This information acquisition department uses web scraping technology to collect unstructured text data from external websites or SNS. This information acquisition department inputs the collected text data into a morphological parsing engine. This engine segments the text into a sequence of word tags. This information acquisition department uses a word embedding model to convert each tag into a 300-dimensional dense vector. This information acquisition department calculates the cosine similarity between the company name vector and the negative word vector. This information acquisition department outputs this similarity as a co-occurrence score. This information acquisition department obtains vulnerability information such as general vulnerability identifiers from a cybersecurity information database. This information acquisition department compares the obtained information with the system structure information of franchise stores. This information acquisition unit calculates a vulnerability severity score when consistent vulnerabilities exist. This unit then integrates this collected data to generate an integrated feature matrix for each franchise store. This integrated feature matrix is ​​used as input data for inference processing in the subsequent decision-making unit. This structure automates the information collection process. Compared to manual human searches, this structure completes comprehensive information collection in milliseconds. This structure brings high efficiency to data collection and prevents omissions. This technology can be applied to supply chain risk management systems or automated due diligence systems for investment funds, etc.

[0103] The judgment department inputs all information collected by the information acquisition department into the generative AI. For example, regarding financial information, the generative AI automatically reads financial statements from the past three periods, automatically calculates sales, profit, equity ratio, whether there is insolvency, the soundness of cash flow, the turnover period of accounts receivable and accounts payable, increases and decreases in fixed assets, the content of investment activities, the shift of short-term and long-term loans, increases and decreases in inventory, and whether there are any tax arrears, etc., and compares them with industry averages or past franchise store data to score the soundness of finances. Regarding all historical documents, the generative AI uses OCR or natural language processing to automatically parse the content, extract the legitimacy of legal person status, the history of changes in representatives, the content recorded in the purpose column, the history of increases and decreases in capital, the history of changes in location, records of dissolution or liquidation, mergers, divisions, business transfers, etc., and compares them with past cases of violations or risk patterns. Regarding personal information protection mechanisms, generative AI analyzes checklists or regulatory documents to automatically determine their suitability to laws and guidelines, the comprehensiveness of information management regulations, the implementation status of employee training, the existence of countermeasures against information leaks, the status of third-party certifications, whether there is entrusted management, and the completeness of personal information management ledgers. Regarding the exclusion of antisocial forces, the presence of risk is determined by comparing the content of pledges or external databases, such as comparing with antisocial force lists or verifying past administrative penalty histories. Regarding past records of illegal transactions and administrative guidance, generative AI retrieves data from external databases or news reports, extracting past administrative penalties or dispute cases, litigation history, the content of administrative guidance, whether there are improvement reports, the implementation status of countermeasures to prevent recurrence, the frequency or severity of administrative guidance, etc., to assess the risk level. Regarding the content recorded on a company's website, generative AI automatically crawls the website and automatically determines the authenticity of the recorded content, whether there is exaggerated advertising, discrepancies with past recorded content, the accuracy of product and service descriptions, whether there are any violations of laws and regulations, whether there are any infringements of third-party rights, and whether there are any false representations or violations of the Goods Representation Act. Regarding industry risks, AI determines whether an industry is high-risk by referring to the overall industry's violation rate or administrative guidance rate, past franchisee dispute rate, expulsion rate from industry associations, industry-specific regulatory risks, and social credit rating. Regarding the suitability of the settlement system integration, technical information or API specifications are analyzed to automatically determine the compatibility or security requirements of the existing settlement system, the strength of encryption methods, the standard compliance of communication protocols, the pass rate of sandbox testing, past system failure history, and the existence of external vulnerability reports. For example, it automatically verifies the recorded content of the API specifications and the specifications of the settlement service side, automatically determines compatibility or security vulnerabilities, and analyzes sandbox test logs to detect any unauthorized communication or abnormal responses.Specifically, this decision-making unit executes a composite inference engine combining a large-scale language model based on the Transformer architecture and a graph neural network. As input data, this unit receives time-series arrangements of financial indicators, labeled embedding vectors of text data, and structured parameters of system linkage specifications. The time-series arrangements of financial indicators are fed into a Long Short-Term Memory (LSTM) network. This network learns the variation patterns of financial data over the past three periods, outputting a continuous score from 0 to 1 representing the probability of future default. The unit constructs a knowledge graph from the relationships between entities extracted from all historical documents. A graph neural network is used to predict hidden links between nodes in this knowledge graph. The network output represents a risk probability distribution indicating whether a network structure similar to past defaulting companies exists. The unit inputs API specifications and sandbox test log data into an anomaly detection model. This model uses an autoencoder to learn the latent space of normal communication patterns. The model calculates the reconstruction error between the input log data and the reconstructed data. If the reconstruction error exceeds a specified threshold, the model outputs an anomalous communication flag. The unit then inputs the output scores from these models into a fully connected layer. This fully connected layer outputs the final overall fitness score and the contribution weights of each risk factor. The system then sends this overall fitness score to the subsequent mapping and notification departments. This structure eliminates existing review processes that rely on human intuition or experience. This structure mathematically evaluates the correlations of nonlinear characteristic quantities in high-dimensional space. This structure brings about a significant improvement in review accuracy and a reduction in review time. This technology can be applied to credit card affiliate review systems or cryptocurrency exchange account opening review systems, etc.

[0104] Specific examples of prompts that generative AI can read include: "Analyze the following financial statement data and score its financial soundness compared to the industry average," "Extract the legitimacy of legal status and past risk factors from the contents of all the following historical certificates and determine the risk level," "Compare the following regulations on personal information protection systems with legal guidelines and assess their suitability," "Scrape the content of the following franchise store websites and determine whether there are any false records or exaggerated advertisements," "Compare the following API specifications and settlement system requirements and determine the suitability of the linkage," and "Analyze the following sandbox test logs and detect any abnormal responses and unauthorized communications." Specifically, this judgment department has a dynamic prompt generation module. This module automatically selects the best prompt template based on the type and amount of input data. This module embeds structured data into the selected template to generate input text for the natural language processing model. The system receives numerical vectors such as sales revenue and operating profit, as well as text strings from specified documents, as input data. The system inserts this data into placeholders in the prompt text. This judgment department then inputs the generated prompt text into a large-scale language model. This model uses a self-attention mechanism to calculate the dependencies between words within the prompt. As output, the model generates a response text for the requested task, along with probability scores from 0 to 1 representing the confidence level of each decision item. For example, the model outputs a numerical value of 0.85 for financial soundness and a label of "no detection" for abnormal communication. The system then passes these outputs to the subsequent scoring module. This design makes the instructions for the model explicit and suppresses bias inference. It balances flexible task definition based on natural language with rigorous numerical output. This design improves the stability of inference results. This technology can be applied to automated response systems and document summarization systems using conversational agents.

[0105] The judgment department scores each item, performs a comprehensive suitability assessment, generates detailed scoring results along with supporting data, AI-based summary comments, comparisons with similar past cases, etc., and sends them to the results display department on the terminal side. The notification department notifies users in real time of the judgment results and reasons for unsuitability, risk alerts, requests for additional information, feedback on AI learning results, and emergency notifications in the event of a cybersecurity incident. For example, unsuitability is determined in cases such as financial soundness below a certain standard, a history of significant administrative guidance, discovery of false information on a website, inadequate personal information protection mechanisms, detection of antisocial forces risks, high industry risks, technical problems in settlement system linkage, failure to pass sandbox testing, or external vulnerability reports. Reasons for these unsuitability are automatically generated and notified to the terminal side. Specifically, this judgment department executes a comprehensive scoring algorithm based on ensemble learning. This judgment department receives vectors of individual scores output from each inference model as input. This judgment department uses a random forest algorithm to weighted sum these individual scores. This judgment department outputs the final comprehensive suitability score as an integer value from 0 to 100. This decision-making unit generates an unsuitability flag when the overall fitness score is lower than a preset threshold. This decision-making unit identifies features that are the main factors contributing to the score reduction by tracking the branching conditions of the decision tree. Based on the identified features, this decision-making unit generates natural language text data explaining the reasons for unsuitability. This notification unit receives the generated text data, the overall fitness score, and the unsuitability flag. This notification unit encapsulates this data into a structured message payload. This notification unit uses a bidirectional communication protocol to asynchronously push and distribute this message payload to the result display unit on the terminal side. This configuration prevents the decision result from being treated as a black box. This configuration logically and transparently demonstrates the basis for the decision. This configuration provides technical benefits such as supporting user decision-making and improving system reliability. This technology can be applied to providing diagnostic evidence in medical diagnostic support systems and generating hazard prediction warnings in autonomous driving systems, etc.

[0106] The Mapping Department automatically categorizes franchisees based on various attributes, including the appropriate judgment and risk assessment results from the Judgment Department, the franchisee's industry, region, transaction scale, risk level, settlement system linkage status, presence of administrative guidance records, intellectual property ownership status, number of employees, capital size, years of establishment, cybersecurity risk level, and past system failure history. It generates a franchisee map using various visualization techniques such as graph structures, maps, heatmaps, network diagrams, tree diagrams, timelines, risk heatmaps, and cybersecurity risk maps. The results are displayed on the terminal side, allowing department heads to refer to them in real time. For example, the risk management manager can use the heatmap to identify the distribution of franchisees in high-risk industries, the sales manager can use the map to identify the distribution of franchisees in newly developed areas, the legal manager can use the timeline to track the time series of franchisees with administrative guidance records, and the system manager can use the risk map to identify franchisees with high cybersecurity risks. Specifically, this Mapping Department drives a multi-dimensional data visualization engine. It receives a multi-dimensional vector group containing the attribute data of franchisees output from the Judgment Department and their risk scores. The Mapping Department executes principal component analysis algorithms. This mapping unit reduces and compresses the high-dimensional feature space into a two- or three-dimensional visualization space. Based on the compressed coordinate data, it generates a graph network structure with each franchise store as a node and transaction relationships and similarities between franchise stores as edges. The unit proportionalizes node size to transaction scale and assigns node colors as gradients to risk scores. It outputs the generated graph network structure as vector graphics data and sends this data to the results display unit on the terminal. This display unit uses hardware acceleration from a parallel processing processor to render a graph with tens of thousands of nodes without latency. This structure transforms massive amounts of text and numerical data into intuitive visual representations. It enables users to instantly identify complex data relationships and outliers. This structure brings about increased efficiency in data analysis and faster decision-making. This technology can be applied to network traffic monitoring systems, visualization of infectious disease spread simulations, and more.

[0107] The AI ​​Learning Management Department continuously learns from the judgment results made by the Judgment Department in the past, new information acquired from external sources, whether disputes occurred after actual transactions, whether administrative guidance was issued, the bankruptcy and merger of franchise stores, growth performance, whether there were any cyberattacks or system failures, etc., to improve the judgment accuracy of generative AI. For example, if a franchise store judged as unsuitable in the past causes problems later, this case is imported as learning data and reflected in future judgment criteria. In the event of new legal amendments or changes in industry guidelines, the AI ​​Learning Management Department also automatically acquires information and reflects it in the judgment department's algorithm. Specifically, this AI Learning Management Department executes a hybrid learning pipeline that combines online learning and batch learning. This AI Learning Management Department collects log data of disputes and system failures that occurred after actual transactions as correct answer labels. This AI Learning Management Department pairs the input feature vectors from past judgments with these correct answer labels to construct a new learning dataset. This AI Learning Management Department inputs this learning dataset into a neural network model. This model calculates the error between the predicted value and the correct answer label. This AI learning management system uses stochastic gradient descent to fine-tune the model's weight parameters in a way that minimizes the error function. The system inputs text data containing legal amendments into the natural language processing model. The model extracts new constraints from the text and updates the rule-based evaluation function. The system then applies the updated weight parameters and evaluation function as new decision criteria in the next inference process. This design enables the system to autonomously adapt to environmental changes and the emergence of new risk patterns. This design provides technical advantages in preventing model accuracy degradation and maintaining long-term predictive performance. This technology can be applied to environments where data distribution changes over time, such as spam filtering systems and algorithmic trading systems in financial markets.

[0108] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart device 14. In the smart device 14, the control unit 46A causes the output device 40 to output the result of the specific processing. The microphone 38B acquires voice representing the user's input on the result of the specific processing. The control unit 46A sends the voice data representing the user's input acquired by the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0109] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0110] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0111] The data generation model 58 is obtained by performing deep learning on a neural network. A prompt containing instructions is input to the data generation model 58, along with inference data such as speech data representing speech, text data representing text, and image data representing images (e.g., data from still images or moving images). The data generation model 58 infers the input inference data based on the instructions shown in the prompt and outputs the inference result in one or more data forms, such as speech data, text data, and image data. The data generation model 58 may include, for example, text generation AI, image generation AI, and multimodal generation AI. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization. The specific processing unit 290 performs the aforementioned specific processing while using the data generation model 58. The data generation model 58 may also be a fine-tuned model to output inference results from prompts without instructions; in this case, the data generation model 58 can output inference results from prompts without instructions. The data processing device 12, etc., includes multiple data generation models 58, including AI other than generative AI. AI other than generative AI includes, but is not limited to, linear regression, logistic regression, decision trees, random forests, support vector machines (SVM), k-means clustering, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks (GAN), or Naive Bayes. Furthermore, AI can also act as an AI agent. Additionally, when the processing described above is performed by AI, this processing may be partially or entirely performed by AI, but is not limited to this example. Moreover, processing performed by AI, including generative AI, can be replaced by rule-based processing, and vice versa.

[0112] Furthermore, the processing performed by the data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart device 14, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart device 14. In addition, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart device 14 or external devices, and the smart device 14 acquires or collects information required for processing from the data processing device 12 or external devices.

[0113] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart device 14.

[0114] [Second Implementation] Figure 3An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0115] like Figure 3 As shown, the data processing system 210 includes a data processing device 12 and smart glasses 214. One example of the data processing device 12 is a server.

[0116] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0117] The smart glasses 214 include a computer 36, a microphone 238, a speaker 240, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, and communication I / F 44 are also connected to the bus 52.

[0118] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0119] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0120] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0121] Figure 4An example of the main functions of the data processing device 12 and the smart glasses 214 is shown. Figure 4 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0122] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0123] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290. The specific processing unit 290 can use the emotion-specific model 59 to infer the user's emotions and perform specific processing based on the user's emotions. The emotion inference function using the emotion-specific model 59 (emotion-specific function) includes various inferences and predictions related to the user's emotions, but is not limited to this example. Furthermore, emotion inference and prediction also include, for example, emotion analysis (analysis).

[0124] In the smart glasses 214, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating as a control unit 46A based on the receive / output program 60 executed on the RAM 48. Furthermore, the smart glasses 214 may also have the same data generation model and emotion-specific model as the data generation model 58 and the emotion-specific model 59, and use these models to perform the same processing as the specific processing unit 290.

[0125] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the smart glasses 214. In the following description, the data processing device 12 will be referred to as the "server" and the smart glasses 214 will be referred to as the "terminal".

[0126] (Example 1) Since the specific processing flow is the same as that in Embodiment 1 described in the first embodiment above, the description is omitted.

[0127] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0128] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the smart glasses 214. In the smart glasses 214, the control unit 46A causes the speaker 240 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0129] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the smart glasses 214, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the smart glasses 214. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the smart glasses 214 or external devices, and the smart glasses 214 acquires or collects information required for processing from the data processing device 12 or external devices.

[0130] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the smart glasses 214.

[0131] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0132] like Figure 5 As shown, the data processing system 310 includes a data processing device 12 and a head-mounted terminal 314. An example of the data processing device 12 is a server.

[0133] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0134] The head-mounted terminal 314 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a display 343. The computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, display 343, and communication I / F 44 are also connected to the bus 52.

[0135] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0136] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0137] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0138] Figure 6 An example of the main functions of the data processing device 12 and the head-mounted terminal 314 is shown. Figure 6 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0139] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0140] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290.

[0141] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0142] The memory 50 stores a receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on the RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on the RAM 48.

[0143] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the head-mounted terminal 314. In the following description, the data processing device 12 will be referred to as the "server" and the head-mounted terminal 314 will be referred to as the "terminal".

[0144] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0145] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0146] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the head-mounted terminal 314. In the head-mounted terminal 314, the control unit 46A causes the speaker 240 and the display 343 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user's input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0147] Furthermore, the processing performed by the aforementioned data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the head-mounted terminal 314, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the head-mounted terminal 314. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the head-mounted terminal 314 or external devices, and the head-mounted terminal 314 acquires or collects information required for processing from the data processing device 12 or external devices.

[0148] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the head-mounted terminal 314.

[0149] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0150] like Figure 7 As shown, the data processing system 410 includes a data processing device 12 and a robot 414. An example of the data processing device 12 is a server.

[0151] The data processing apparatus 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" as understood in this disclosure. The computer 22 includes a processor 28, RAM 30, and memory 32. The processor 28, RAM 30, and memory 32 are connected to a bus 34. Furthermore, the database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. An example of the network 54 includes a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0152] Robot 414 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a control object 443. Computer 36 includes a processor 46, RAM 48, and memory 50. The processor 46, RAM 48, and memory 50 are connected to a bus 52. Furthermore, the microphone 238, speaker 240, camera 42, control object 443, and communication I / F 44 are also connected to the bus 52.

[0153] Microphone 238 receives instructions from user 20 by receiving the sounds emitted by user 20. Microphone 238 captures the sounds emitted by user 20, converts the captured sounds into speech data, and outputs it to processor 46. Speaker 240 outputs sound according to the instructions from processor 46.

[0154] Camera 42 is a small digital camera equipped with an optical system such as a lens, aperture and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, used to capture the area around user 20 (e.g., the shooting range defined by an angle of view equivalent to the field of vision of an average healthy person).

[0155] Communication I / F 44 is connected to network 54. Communication I / F 44 and 26 are responsible for the transmission and reception of various information between processor 46 and processor 28 via network 54. The transmission and reception of various information between processor 46 and processor 28 using communication I / F 44 and 26 is performed in a secure state.

[0156] The controlled object 443 includes a display device, LEDs for the eyes, and motors for driving the arms, hands, and feet. The posture and movements of the robot 414 are controlled by controlling the motors for the arms, hands, and feet. Some of the robot 414's emotions can be expressed by controlling these motors. In addition, facial expressions of the robot 414 can also be expressed by controlling the illumination state of the LEDs for the robot 414's eyes.

[0157] Figure 8 An example of the main functions of the data processing device 12 and the robot 414 is shown. For example... Figure 8 As shown, in the data processing device 12, specific processing is performed by the processor 28. The memory 32 stores a specific processing program 56.

[0158] The specific processing program 56 is an example of a "program" as disclosed herein. The processor 28 reads the specific processing program 56 from the memory 32 and executes the read specific processing program 56 on the RAM 30. Specific processing is implemented by the processor 28 operating as a specific processing unit 290 according to the specific processing program 56 executed on the RAM 30.

[0159] The memory 32 stores a data generation model 58 and an emotion-specific model 59. The data generation model 58 and the emotion-specific model 59 are used by the specific processing unit 290.

[0160] In robot 414, the processor 46 performs the receive / output processing. The memory 50 stores the receive / output program 60. The processor 46 reads the receive / output program 60 from the memory 50 and executes the read receive / output program 60 on RAM 48. The receive / output processing is implemented by the processor 46 operating the control unit 46A based on the receive / output program 60 executed on RAM 48.

[0161] Next, the specific processing performed by the specific processing unit 290 of the data processing device 12 will be described. Each part of the system described below is implemented by the data processing device 12 and the robot 414. In the following description, the data processing device 12 will be referred to as the "server" and the robot 414 will be referred to as the "terminal".

[0162] (Example 1) Since the specific processing flow is the same as that in Example 1 described in the first embodiment above, the description is omitted.

[0163] (Application Example 1) Since the specific processing flow is the same as that in Application Example 1 described in the first embodiment above, the description is omitted.

[0164] The specific processing unit 290 sends the result of the specific processing obtained in Embodiment 1 or Application Example 1 described above to the robot 414. In the robot 414, the control unit 46A causes the speaker 240 and the controlled object 443 to output the result of the specific processing. The microphone 238 acquires voice representing the user's input on the specific processing result. The control unit 46A sends the voice data representing the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[0165] Furthermore, the processing performed by the data processing system 10 is executed by the specific processing unit 290 of the data processing device 12 or the control unit 46A of the robot 414, but it can also be executed jointly by the specific processing unit 290 of the data processing device 12 and the control unit 46A of the robot 414. Additionally, the specific processing unit 290 of the data processing device 12 acquires or collects information required for processing from the robot 414 or external devices, and the robot 414 acquires or collects information required for processing from the data processing device 12 or external devices.

[0166] In the above embodiments, examples of specific processing performed by the data processing device 12 are given, but the technology disclosed herein is not limited to this, and specific processing may also be performed by the robot 414.

[0167] Furthermore, the emotion-specific model 59, serving as an emotion engine, can determine the user's emotion based on a specific mapping. Specifically, the emotion-specific model 59 can determine the user's emotion based on an emotion graph that serves as a specific mapping (see...). Figure 9 The robot's emotions can be determined by the emotion-specific model 59. In addition, the emotion-specific model 59 can also determine the robot's emotions in the same way, and the specific processing unit 290 can also perform specific processing using the robot's emotions.

[0168] Figure 9 This is a diagram representing an emotion map 400 that maps various emotions. In the emotion map 400, emotions are arranged radially from the center in concentric circles. The closer to the center of the concentric circles, the more primitive the emotion is. Further out on the concentric circles, emotions are arranged representing states or actions arising from mood. Emotion is a concept that includes both feelings and mental states. To the left of the concentric circles, emotions generated by reactions occurring in the brain are arranged roughly. To the right of the concentric circles, emotions guided by situational judgments are arranged roughly. Above and below the concentric circles, emotions generated by reactions occurring in the brain and guided by situational judgments are arranged roughly. Furthermore, the emotion of "pleasure" is arranged above the concentric circles, and the emotion of "unpleasantness" is arranged below. Thus, in the emotion map 400, various emotions are mapped according to the structure of emotion generation, while easily generated emotions are mapped nearby.

[0169] These emotions are distributed at the 3 o'clock position on the Emotion Chart 400, and usually fluctuate between peace and unease. In the right half of the Emotion Chart 400, because situational awareness is more dominant than internal feelings, it gives a sense of calm.

[0170] The inner side of the emotion diagram 400 represents the mind, and the outer side of the emotion diagram 400 represents actions. Therefore, the further you go to the outer side of the emotion diagram 400, the more the emotion can be seen (manifested in actions).

[0171] Here, human emotions are based on a balance of various factors such as posture and blood sugar levels. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. In robots, cars, and motorcycles, emotions can also be created based on a balance of factors such as posture and remaining battery power. When these balances deviate from the ideal, it indicates unhappiness; when they approach the ideal, it indicates pleasure. Emotion maps can be generated, for example, based on Dr. Mitsuyoshi's emotion map (Research on Speech Emotion Recognition and Brain Physiological Signal Analysis Systems for Emotion, Tokushima University, Doctoral Dissertation: https: / / ci.nii.ac.jp / naid / 500000375379). In the left half of the emotion map, emotions belonging to the "Reaction" domain, where sensation is dominant, are arranged. Furthermore, in the right half of the emotion map, emotions belonging to the "Situation" domain, where situational cognition is dominant, are arranged.

[0172] The emotion map defines two types of emotions that promote learning. One is a negative emotion located near the middle of "repentance" or "reflection" on the situation side. That is, when the robot experiences negative emotions such as "I never want to feel this way again" or "I never want to be scolded again." The other is a positive emotion located near "desire" on the response side. That is, when the robot experiences positive feelings such as "wanting more" or "wanting to know more."

[0173] The emotion-specific model 59 feeds user input into a pre-learned neural network to obtain emotion values ​​representing each emotion shown in the emotion graph 400, and determines the user's emotion. This neural network is pre-learned based on multiple learning data sets that combine user input with emotion values ​​representing each emotion shown in the emotion graph 400. Furthermore, this neural network is learned to... Figure 10 As shown in sentiment graph 900, nearby sentiment values ​​are similar to each other. Figure 10 Examples show how emotions such as "peace of mind," "stability," and "reassurance" can result in similar emotional values.

[0174] The above description primarily focuses on the functions of the data processing...

Claims

1. A customer information management system, characterized in that, It includes an information acquisition department, a judgment department, and a mapping department. The Information Acquisition Department automatically collects multi-dimensional corporate information from multiple databases inside and outside the company and external information sources, including the company name, representative name, financial information for multiple periods, all historical certificates, personal information protection system, past business performance, administrative guidance records, website content, alliance information, business plans, and regional information of new customer candidates. The judgment department uses generative AI to analyze various information obtained, and scores and determines suitability from multiple perspectives, such as financial soundness, legal status, suitability of personal information management system, reliability of past business performance, whether there are administrative guidance records, authenticity of website content, appropriateness of alliance partners, suitability of business plan, and whether there is overlap with existing customers. Based on the judgment results, it automatically outputs whether registration is possible and the reasons for unsuitability. as well as The mapping department categorizes registered customer information according to attributes such as region, industry, transaction size, and risk level, visualizes it as a corporate map, and manages past registration history and assessments through a database. In addition, the determination unit continuously learns from past determination results and external information to improve determination accuracy.

2. The system as described in claim 1, characterized in that, The information acquisition department comprehensively collects customer information managed by multiple departments within the company, including customer management, legal compliance, accounting and finance, and sales. It also collaborates with external credit investigation agencies or government agencies to obtain publicly available information, as well as information from the company's website on the internet. The decision-making department uses generative AI to comprehensively analyze the data obtained from these diverse information sources, and unifies and automatically executes the investigations and reviews that were previously conducted separately by each department.

3. The system as described in claim 1, characterized in that, The mapping unit combines the appropriate judgment results and risk assessment results made by the judgment unit, the company's region, industry, transaction scale, past administrative guidance records, alliance member attributes, business plan content, and other attribute information to dynamically visualize the relationship and risk distribution of customer candidates and existing customers as a corporate map. This allows various departments within the company to refer to this map information in real time. The decision-making unit continuously improves the accuracy of the enterprise map and the reliability of risk assessment by using generative AI to learn new external information and past decision history.

Images