Data processing system

Abstract

The present application relates to a kind of contact center systems, it includes customer information management department, operator information management department and generative AI processing part, a variety of information such as the consultation history and attribute of each customer, the skill and operation condition of operator are managed comprehensively, the analysis of consultation content is carried out using generative AI, emotion recognition, the self-generation of optimal coping strategy and response text is automatically generated.And, customer satisfaction and feedback from operator are applied to self-learning, continuously optimize model precision and response content, and maintain coping quality autonomously through self-correction function when problem occurs.Thereby, while realizing customer satisfaction improvement, business efficiency and operator load reduction, self-evolution type contact center application that can be flexibly adapted to all industries and business formats is realized.

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 present invention aims to comprehensively address several challenges faced by enterprises operating contact centers. These challenges include: decreased customer satisfaction due to the inability of traditional rule-based response systems to handle complex and emotionally charged customer interactions; high-stress environments and increased turnover rates among operators due to monotonous business and emotional labor; cost-inefficiency deterioration due to increased staffing and training costs caused by lagging business efficiency improvements; and insufficient data utilization due to the failure to fully leverage customer data and past consultation history. Specifically, the invention involves comprehensively managing various information such as each customer's past history and attributes, operator skills and stress levels, etc., using generative AI to grasp customer emotions and needs in real time and with high precision. By automatically generating optimal response strategies, operators and AI collaborate to improve business efficiency and customer satisfaction. Furthermore, by possessing self-learning and self-correction capabilities based on feedback loops, the invention flexibly adapts to the customization and market changes of each enterprise, thereby reducing customer service costs, increasing customer satisfaction, and lowering operator turnover. Summary of the Invention

[0005] The means to solve the problems involved in this invention is to construct a contact center system comprising a customer information management department, an operator information management department, and a generative AI processing department. In the customer information management department, the system stores each customer's consultation history, attribute information, past reception results, customer preferences, and channel information in the long term, and can refer to the customer's past history and attribute information in real time when a consultation occurs. In the operator information management department, the system manages each operator's skill set, responsibility history, stress level, and scope of business they can handle, and dynamically allocates and controls tasks based on the operator's operational status and adaptability. In the generative AI processing department, generative AI is used to parse the consultation content from customers using natural language, identify the emotions in the consultation text in real time, and automatically generate the best response strategy while comprehensively referring to the customer's past history and attributes, the operator's skills and stress level, etc. Furthermore, the system continuously performs self-learning using customer satisfaction and feedback information obtained after customer reception, autonomously optimizing model accuracy and response content, adapting to changes in consultation patterns and market environments, and autonomously readjusting model parameters through a self-correction function when problems occur, thereby improving customer satisfaction, increasing business efficiency, and reducing operator turnover. 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) This invention provides a contact center system comprised of a server and terminals, including a customer information management department, an operator information management department, and a generative AI processing department. The server performs the core functions of the customer information management department, operator information management department, and generative AI processing department. The terminals provide an operator user interface and, as needed, a lightweight AI inference module. The following provides a detailed and comprehensive description, including the structure and function of each department, methods for utilizing generative AI, the role division between the server and terminals, and specific application examples. Specifically, this contact center system adopts an architecture that links multiple computing nodes in a distributed computing environment. The server consists of a physical server or a cluster of virtual servers equipped with multi-core processors and large memory. The server implements the customer information management department and the operator information management department as relational databases or NoSQL databases. The server builds the generative AI processing department on a parallel computing cluster using GPUs. The terminal is a personal computer or smart device that uses edge computing technology to execute the lightweight AI inference module. This lightweight AI inference module includes a neural network model with a reduced number of parameters through quantization or knowledge distillation. This system dynamically switches between large-scale inference processing on the server side and low-latency inference processing on the terminal side. Through this architecture, the system achieves a significant improvement in overall system response speed while suppressing network bandwidth consumption. This system is particularly useful in applications requiring real-time performance and high security, such as customer support in financial institutions or appointment centers in medical institutions. This system replaces traditional manual information retrieval with automatic extraction based on similarity calculations in a high-dimensional vector space. As a result, the computation time required for information retrieval is reduced to the millisecond level.

[0040] The Customer Information Management Department, implemented on a server, provides long-term storage and management of each customer's consultation history and attribute information (name, age, gender, contract details, services used, past purchase history, past complaint history, past satisfaction scores, channels used, reception history, questionnaire results, etc.). For example, in the financial industry, it comprehensively manages past transaction history and any instances of improper use; in the medical industry, it comprehensively manages past medical history and appointment history; in the retail industry, it comprehensively manages purchase and return history; and in the telecommunications industry, it comprehensively manages contract packages and past fault handling history—information required by each industry. When a customer makes a consultation, the consultation content is sent from the terminal to the server. The Customer Information Management Department immediately retrieves the customer's entire history and provides it to the Generative AI Processing Department. Specifically, this Customer Information Management Department structures customer information as multi-dimensional time-series data and stores it in a database. It uses a natural language processing model to convert each customer's attribute and historical information into fixed-length numerical vectors. The transformed numerical vectors are then indexed and saved in a vector database. This customer information management department encodes the text data of the consultation content received from the terminal into numerical vectors. It then calculates the cosine similarity between the numerical vectors of the consultation content and the numerical vectors of customer information in the database. This department rapidly extracts historical data with similarity exceeding a predetermined threshold. Unlike previous keyword-based retrieval methods, this extraction process uses distance calculation in a high-dimensional space to evaluate semantic relevance—a less conventional approach. The department generates a multidimensional tensor that rearranges the extracted historical data in chronological order. This multidimensional tensor is then provided as input to the generative AI processing unit. Through this structure, the system achieves the technical effect of instantly identifying context-appropriate information from massive amounts of customer data. This system can be applied in call centers for telemarketing, enabling personalized product recommendations based on customers' past purchasing tendencies. This system reduces the computational load of data retrieval, lowering the CPU load on the server.

[0041] The Operator Information Management Department is also implemented on the server, uniformly managing each operator's skill set (e.g., "responding to credit card misuse," "medical consultation," "delivery arrangements," "contract changes," "fault handling," etc.), responsibility history (what kinds of consultations were handled in the past and how many), stress level (number of recent cases handled, customer feedback, rest status, AI-based sentiment analysis results, etc.), scope of business they can handle (e.g., "medical field only," "can handle both finance and retail," etc.), and operational status (currently handling cases, standby, overtime status, etc.). For example, it can monitor in real time information such as Operator A's expertise in the medical field, Operator B's ability to handle both finance and retail, and Operator C's high stress level as a newcomer. Specifically, this Operator Information Management Department maintains each operator's status as a dynamically updated state vector. It represents each operator's skill set as a multidimensional array with one-hot encoding. It collects operators' voice data and typing speed in real time. The department preprocesses the collected raw data and inputs it into a recurrent neural network. This recurrent neural network outputs a continuous score from 0.0 to 1.0 representing the operator's fatigue or stress level from the raw time-series data. When the stress level score exceeds 0.8, the operator information management department automatically changes the operator's operational status flag to recommend rest. This department integrates the state vectors of each operator to generate a resource distribution matrix for the entire operator pool. Using this resource distribution matrix, the system determines the optimal operator allocation in subsequent routing processes. This process differs from traditional business management, which relies on visual judgment by human managers; it is an objective and quantitative algorithmic process based on biometric information and operational logs. Through this structure, the system achieves the technical effect of preventing operator overload and optimizing the overall system throughput. This system is applied in large-scale technical support centers to automatically assign complex inquiries to operators with high skill scores.

[0042] The generative AI processing unit is implemented on a server, using generative AI as an essential element. Generative AI is composed of multiple AI technologies, such as large-scale language models, natural language processing models, sentiment recognition AI, predictive modeling AI, and clustering AI. Generative AI uses natural language parsing to analyze the consultation content sent from the terminal, accurately identifying the sentiment in the consultation text. For example, it identifies "confused" for the consultation "I'm upset that the goods haven't arrived," and "angry" for the consultation "Please don't make me say the same thing several times." In sentiment recognition, it combines external sentiment recognition engines or independently developed deep learning models, and also uses speech sentiment analysis in the case of voice input. Specifically, this generative AI processing unit has a large-scale language model based on the Transformer architecture as its core. This generative AI processing unit tokenizes the text data received from the terminal, for example, converting it into 512-dimensional word embedding vectors. This generative AI processing unit adds positional encoding to the word embedding vectors and inputs them into a multi-layered self-attention mechanism. This generative AI processing unit extracts the semantic representation vector of the entire sentence from the output of the self-attention mechanism. This generative AI processing unit inputs semantic representation vectors into a fully connected layer and a Softmax function. As the output of the Softmax function, this generative AI processing unit outputs probability distributions for various emotion categories, such as anger, sadness, joy, and neutrality. For example, for the input of damaged goods, this generative AI processing unit outputs a probability of 0.85 for anger and 0.10 for sadness. In the case of voice input, this generative AI processing unit converts the voice waveform into acoustic feature tensors such as Mel-frequency cepstral coefficients. This generative AI processing unit inputs the acoustic feature tensors into a convolutional neural network and outputs an emotion score based on the pitch and intonation of the voice. This generative AI processing unit uses ensemble learning to integrate the text-based emotion probability distribution and the voice-based emotion score. This generative AI processing unit then passes the final integrated emotion label to the subsequent response generation module. Through this structure, this system effectively detects potential customer dissatisfaction early and prevents complaints from escalating. This system can be applied to use cases where emergency roadside assistance is needed, such as detecting the panic state of customers and automatically prioritizing them.

[0043] Generative AI comprehensively considers customer history and attributes, operator skills, and stress levels to automatically generate optimal response strategies. For example, for customers with repeated similar complaints, it generates a faster solution or a special apology based on previous responses. In the medical field, it automatically generates appropriate treatment guides and precautions based on patients' past medical history and allergy information. In the financial field, it generates rapid self-verification steps and compensation guides based on past fraud history and personal verification information. In the retail field, it automates return and exchange procedures and offers discounts for future purchases based on past purchase and return history. In the communications field, it combines group notifications and individual responses when a failure occurs, such as generating individual messages like "Currently experiencing a failure. Estimated recovery time is 0:00. We apologize for the inconvenience." Specifically, this generative AI combines customer history tensors obtained from the customer information management department, state vectors obtained from the operator information management department, and labels from emotion recognition results to generate a unified context vector. This generative AI inputs this unified context vector as part of the prompt into a large-scale language model. This large-scale language model generates tags autoregressively based on the integrated context vector of the input. As output data, this large-scale language model generates structured data formats representing response steps and natural language response text prompting customers. For example, as input, this large-scale language model receives three past complaints, current emotion of anger, and operator skill level of "moderate." As output, this large-scale language model generates response strategies such as "Recommend escalating to a higher level" and the response text "I'm very sorry for the great trouble you've caused; please immediately assign a different person to handle this." To evaluate the appropriateness of the generated response strategies, this generative AI uses a reinforcement learning-based reward model to score the generated text. If the score falls below a predetermined threshold, the generative AI regenerates the response text using different temperature parameters. This process replaces the human operator's mental process of searching manuals and considering responses with a probabilistic tag generation algorithm in a high-dimensional space. Through this structure, the system achieves the technical effect of executing response strategy decisions with complex conditional branches in milliseconds, minimizing response latency. This system is applied in airline reservation centers to a use case where alternative flight proposals and compensation guides are personalized and automatically generated for each customer in the event of flight cancellation.

[0044] The generative AI provides real-time prompts to the operator terminal, including generated response strategies, response texts, supplementary proposals, precautions, FAQ links, and relevant manuals. For example, the operator screen may display specific instructions such as, "This customer has made the same inquiry three times in the past. Since an apology and discount coupon were offered last time for the delivery delay, we recommend adding an apology and a detailed explanation of the delivery status this time." Furthermore, the AI-generated response text can be sent directly via chat or automatically responded via speech synthesis. While referring to the AI's proposals, the operator can fine-tune the response content based on their own judgment as needed. Specifically, this generative AI packages the generated response strategy and response text data along with display metadata and sends it to the operator terminal. The operator terminal parses the received data packet and dynamically renders it in a designated area on the graphical user interface. The operator terminal automatically applies highlighting or color coding based on importance to the response text data. The operator terminal has a text-to-speech conversion engine. The operator terminal converts the generated response text into speech waveform data and executes an automatic response based on speech synthesis. When an operator edits the response text, this operator terminal uses algorithms such as Levinstein distance to calculate the difference between the text before and after editing. The operator terminal asynchronously sends the calculated difference data to the server as feedback information. This processing is not simply information display, but a technical process of real-time parsing of the operator's operation log and achieving coordinated actions between the system and the human. Through this structure, the system significantly reduces the operator's input load and physically prevents typos and wizard errors. This system can be applied to use cases where operators can send AI-generated translated response text with a single click in multilingual help desks. To optimize communication traffic, this system only compresses and sends the difference data.

[0045] After a customer is served, customer satisfaction scores (e.g., "very satisfied," "somewhat dissatisfied," etc.) and operator feedback (e.g., "the AI ​​proposal is accurate," "the proposal doesn't fit the situation") are sent from the terminal to the server. The generative AI processing unit accumulates and analyzes this data. The generative AI uses this feedback to continuously perform self-learning, optimizing model parameters and response generation algorithms. For example, if satisfaction is low in a specific response pattern, the response text and steps for that pattern are automatically corrected and improved. Conversely, patterns with high satisfaction are actively promoted to other operators. Furthermore, when operators are under high stress, the AI ​​can automatically adjust the number of responses or recommend rest periods. Specifically, the server accumulates the received customer satisfaction scores and operator edit difference data as teacher data. The generative AI processing unit uses this accumulated teacher data to periodically fine-tune a large-scale language model. In this fine-tuning, the generative AI processing unit employs parameter efficiency techniques such as low-rank adaptation. With the original model's weight matrix fixed, the generative AI processing unit updates the added low-rank matrix only using gradient descent. This generative AI processing unit uses cross-entropy error as the loss function to minimize the error between the generated response and the operator's actual corrected response. It executes a proximal policy optimization algorithm that uses customer satisfaction scores as a reward signal. This unit reduces the probability of generating low-satisfaction response patterns and increases the probability of generating high-satisfaction patterns. The processing explicitly defines the learning process of AI, which is prone to being black-boxed, as specific mathematical operations involving error backpropagation and weight updates. Through this structure, the system leverages its technical advantage of increasingly self-regulating response accuracy over longer application periods, preventing model obsolescence. This system is applied to customer support in the seasonally fluctuating apparel industry, enabling the model to quickly adapt to the latest fashion trends and changing customer preferences. The system utilizes the idle resources of the GPU cluster to perform learning processing in the background, eliminating the impact on inference processing in the production environment.

[0046] When generative AI detects new consultation patterns or changes in the market environment, it autonomously generates response strategies using clustering or pattern recognition techniques and rapidly reflects these strategies across the entire system. For example, if inquiries about "not knowing how to use the new service" increase sharply shortly after a new service is launched, the AI ​​automatically generates FAQs or manuals and displays them instantly on the operator's screen. When laws and regulations change, it automatically extracts relevant inquiries and generates response documents based on the latest legal information. During disasters, it can also automatically send safety confirmations or support information to customers in affected areas. Specifically, this generative AI applies unsupervised clustering algorithms in real-time to the embedded vector groups of daily input consultation texts. When a new dense region of vector groups forms in a high-dimensional vector space at a distance greater than a certain Euclidean distance from existing clusters, this generative AI detects it as a new consultation pattern. When a new pattern is detected, the generative AI extracts multiple text data points closest to the center point of that vector group. The extracted text data is then input into a summarization model, automatically extracting the core themes of the new pattern. This generative AI combines extracted topics with the latest information from external knowledge bases to automatically generate tagged data for new answer entries. This generative AI then submits the generated tagged data to the system's shared database in real time, invalidating the cache on all operator terminals and updating the display. This processing, through density calculations in vector space and automatic text generation algorithms, completely replaces the previous process of human managers manually analyzing trends and creating manuals. Through this architecture, the system drastically reduces the system's adaptation time to unknown events from days to minutes. This system can be applied to IT service desks experiencing large-scale system failures, instantly detecting a sharp increase in failure reports and automatically sending temporary avoidance strategies to all operators.

[0047] From a security and privacy perspective, customer information and reception history are strictly managed on the server, with only the minimum necessary information sent to the terminal. For example, the terminal can be configured to display response strategies and necessary historical information in a way that doesn't specify a particular individual. Furthermore, some inference processing and user interface aspects of generative AI can be executed on the terminal, designed to perform minimal response generation and history referencing even offline or with network latency. Specifically, the server performs data encryption using advanced encryption standards on the customer information and reception history database. Before sending data to the terminal, the server dynamically masks personal information such as names or credit card numbers contained in the text data using regular expression matching and named entity recognition models. The server only sends masked text data to the terminal via a secure communication protocol. The terminal maintains a lightweight language model in its local memory. If the communication latency with the server exceeds a predetermined threshold, or if the network connection is interrupted, the terminal automatically switches to inference mode using the local lightweight model. The terminal's lightweight model generates a provisional response text from the input text data within the local environment. When network connectivity is restored, this terminal synchronizes the locally generated and sent response history to the server. This process goes beyond simple data confidentiality; it employs an advanced distributed processing architecture that dynamically allocates computing resources between the cloud and edge based on network conditions. Through this architecture, the system mathematically reduces the risk of information leakage while maintaining system availability during communication failures. This system is applicable to use cases where temporary call centers located in remote areas with unstable communication infrastructure or disaster sites provide uninterrupted customer support.

[0048] For example, in the financial sector, when a customer inquires about "my card being misused," generative AI immediately refers to past usage history and personal confirmation information to automatically generate a response such as, "To confirm your recent usage history, please provide your date of birth and registered phone number for verification. Details of the reimbursement procedure are as follows," and then displays this message to the operator. In the medical field, when a patient inquires about "changing an appointment," generative AI refers to past appointment history and medical history to generate a response such as, "Your current appointment is on [Date] at [Time]. If you can provide your preferred date and time, we will check availability and guide the doctor to make the change if necessary." In the retail sector, when a customer inquires about "returning an item," generative AI refers to past purchase and return history to generate a personalized response such as, "The item you purchased was shipped on [Date]. The return procedure can be completed using the following steps. Shipping was free last time, but the conditions are different this time, please note." In the telecommunications field, when a customer inquires about "not being able to connect to the internet," generative AI, referencing fault information and contract packages, provides real-time information such as, "Currently, a fault has occurred in region 0. The estimated restoration time is 0:00. We apologize for any inconvenience caused." Specifically, this generative AI internally possesses multiple expert models that have learned domain-specific knowledge for various industries. Based on the vector representation of the input inquiry text, this generative AI dynamically selects the optimal expert model via a routing network. When selecting an expert model in the financial field, this generative AI receives customer identifiers, time-series tensors of transaction history, and fraud detection scores as input data. The financial expert model outputs a structured data record confirming recent usage history, setting the self-verification requirement flag to true. When selecting an expert model in the medical field, this generative AI receives patient identifiers, text vectors of electronic medical records, and a matrix of appointment schedules as input data. The medical expert model outputs a list of changeable appointment times, and the current appointment information as the generated text. When selecting an expert model in the retail field, this generative AI receives order identifiers, purchase history vectors, and a set of return policy rules as input data. This retail expert model sets the return condition flag to paid as output data, and outputs the data as generated text indicating that the purchased goods are as follows. This generative AI, using a selected expert model in the communications field, receives line identifiers, network device life / death monitoring logs, and location coordinates as input data. This communications expert model sets the failure flag to true as output data, and outputs the data as generated text indicating that the user is currently in the relevant region. This processing does not use a single, massive model to handle everything, but rather employs a non-traditional architecture that concentrates computational resources on specific sub-networks based on the task. Through this structure, the system achieves the technical effect of significantly reducing computational costs and power consumption while maintaining high inference accuracy across multiple domains.This system can be applied to the integrated contact center of a diversified enterprise that conducts multiple businesses, using a single system platform to handle all business inquiries.

[0049] Furthermore, generative AI can also be applied to operator education and training. For example, when a new operator is responding to a situation, generative AI can suggest past successful responses or display suggestions in real time. It can also be equipped with functions that automatically evaluate the operator's service and provide feedback and improvement points. Specifically, this generative AI has a dialogue agent module that acts as a virtual customer in a simulated environment. This dialogue agent module reads various emotional parameters and speech patterns extracted from past actual consultation logs as initial prompts. This dialogue agent module receives the new operator's input text or speech recognition text in real time. For the received input, this dialogue agent module generates a natural language response based on set emotional parameters, executing a simulated dialogue session. In the dialogue session, this generative AI uses natural language processing algorithms to quantitatively score the operator's response speed, keyword coverage, and empathy. This generative AI outputs, for example, a multi-faceted evaluation matrix with a solution suggestion speed score of 85 and an empathy score of 40. This generative AI extracts contextually relevant text data from the response logs of high-scoring operators in the past for low-scoring items, displaying improvement suggestions as pop-ups on the screen. This process replaces existing on-site training with full human guidance with data-driven automated evaluation and feedback loops. Through this architecture, the system significantly reduces the time required for operators to become proficient and lowers training costs. This system is suitable for use cases in large-scale call centers with high turnover rates and frequent new hiring, aiming to homogenize training quality and improve service levels.

[0050] Thus, this invention, by implementing the core functions of a customer information management department, an operator information management department, and a generative AI processing department on the server side, and configuring an operator interface and a lightweight AI inference module as needed on the terminal side, achieves real-time understanding of each customer's context and emotions, automatically generates and prompts the best response strategy, and continuously executes a highly advanced contact center system with self-learning, self-correction, and environmental adaptation capabilities. Generative AI plays an indispensable role in all processes, including consultation analysis, emotion recognition, response generation, self-learning, self-correction, environmental adaptation, operator education, trend analysis, automatic FAQ generation, automatic manual updates, simultaneous notification during disasters, and response to changes in laws and regulations. It enables high-precision and flexible customer responses that existing rule-based systems cannot achieve. Specifically, this system adopts a hybrid architecture that links the cloud computing environment with edge devices. The system dynamically distributes inference tasks between a large-scale neural network model on the server side and a lightweight model on the terminal side. The system continuously measures the complexity of input data and network latency through a monitoring module. Based on the measured indicators, the system executes a routing algorithm to determine whether each inference request is processed on the server or the terminal. This system minimizes overall latency and maximizes throughput through dynamic routing. It containerizes the generative AI modules as microservices, using orchestration tools to automatically scale horizontally based on load. This approach fundamentally differs from static, rule-based systems, representing an improvement in computer technology that autonomously optimizes resources in response to fluctuating data traffic. This architecture enables the system to prevent downtime and maintain stable service delivery even in the face of sudden surges in inquiries. The system is suitable as a mission-critical customer response platform operating in environments with unpredictable traffic, such as shortly after a TV shopping broadcast or during large-scale disasters. This system maximizes hardware resource utilization and reduces operating costs.

[0051] (System Composition) The system described in this implementation includes a Customer Information Management Department, an Operator Information Management Department, a Generative AI Processing Department, a Terminal Interface Department, a Feedback Accumulation Department, a Security Management Department, and an External Linkage Department. The Customer Information Management Department stores and manages each customer's consultation history, attribute information, past reception results, purchase history, complaint history, questionnaire results, usage channels, contract content, past satisfaction scores, etc. For example, in the financial sector, it comprehensively manages past transaction history and any instances of improper use; in the medical sector, it comprehensively manages past medical history and appointment history; in the retail sector, it comprehensively manages purchase and return history; and in the telecommunications sector, it comprehensively manages contract packages and fault handling history, among other information required by each industry. When a customer inquires through various channels such as chat, voice calls, or email, the Customer Information Management Department immediately retrieves the customer's entire history and provides it to the Generative AI Processing Department. For example, if there are past consultation histories such as "credit card suspension" or "medical appointment changes," these histories are immediately referenced. Furthermore, as customer attribute information, it also manages age, gender, residential area, family structure, terminal used, and past activity history, and utilizes this information in personalized optimization responses. Specifically, this system functions as an integrated information processing device combining multiple functional modules via a high-speed internal bus or virtual network. The Customer Information Management Department uses a distributed graph database to store customer data. This department constructs a graph structure where customers are nodes and purchase and consultation histories are represented as edges. When receiving a new consultation from a customer, the department uses a graph neural network to analyze the surrounding graph structure of that customer node. As the output of the graph neural network, the department generates a high-dimensional feature vector representing the customer's current status and potential needs. For example, the department receives customer node identifiers, transaction edge groups from the past three months, and recent complaint edges as input data. The department generates a feature vector as output data, containing predicted values ​​for setting a cancellation risk score to 0.85 and a recommended response category to a special discount proposal. The department then forwards the generated feature vector to the generative AI processing department. This process replaces the simple querying and retrieval of relational databases with a non-traditional data processing method based on graph-structured deep learning algorithms. Through this structure, the system not only computationally extracts the explicit content of customer inquiries but also computationally extracts the complex underlying context and relationships, dramatically improving the accuracy of responses. This system is suitable for use cases in customer success departments of subscription services, enabling early detection of signs of contract termination and automatic execution of proactive retention strategies.

[0052] The Operator Information Management Department centrally manages each operator's skill set, responsibility history, stress level, scope of services they can handle, operational status, rest status, past customer reviews, educational history, and qualification information. For example, it monitors in real time information such as Operator A's expertise in the medical field, Operator B's ability to handle both financial and retail clients, Operator C's high stress level as a newcomer, Operator D's high past evaluations, and Operator E's ability to handle nighttime cases. When dynamically allocating operators best suited to the consultation content and client attributes—for example, prioritizing operators with medical skills for consultations requiring medical expertise, and automatically adjusting the workload of operators with high stress levels—the department can also control the process by ensuring that only qualified personnel can handle specific consultations, based on the operator's educational history and qualification information. Specifically, the Operator Information Management Department expands each operator's profile data into a multi-dimensional capability matrix in memory. The department calculates a matching score using an attention mechanism between the feature vector of incoming consultations and the capability matrices of each operator in standby mode. In calculating the matching score, this Operator Information Management Department uses a unique evaluation function that incorporates the operator's current stress level and continuous operating time as penalty factors. The department identifies the operator with the highest overall score calculated by the evaluation function and sends a session connection request to that operator's terminal. For example, the department receives input data such as high consultation difficulty, required medical skills, and customer emotional anger. The department calculates the score for each operator and outputs the best operator identifier D and a matching score of 0.92 as output data. Where specific qualification requirements exist, the department applies a hard constraint filter that forcibly sets the score of operators without qualification information to zero. This process replaces simple polling or rule-based assignment with a real-time solution to a dynamic multivariate optimization problem. Through this architecture, the system achieves the technical effects of preventing operator skill mismatch, reducing average processing time, and improving the first-response rate. This system is applicable to use cases in securities company call centers where reliable routing for operators with specific financial product trading qualifications is performed to ensure compliance with regulations.

[0053] The Generative AI Processing Unit utilizes generative AI as an essential element, combining multiple AI technologies such as natural language processing models, large-scale language models, emotion recognition AI, predictive modeling AI, and clustering AI. It uses natural language parsing to analyze inquiries sent from the terminal interface, accurately identifying the emotions within the text. For example, an inquiry about "being troubled by the undelivered goods" is identified as "confused," while "please don't make me repeat the same thing several times" is identified as "angry." In emotion recognition, it combines external emotion recognition engines or independently developed deep learning models, also using voice emotion analysis in the case of voice input. Generative AI comprehensively considers information such as the customer's past history and attributes, the operator's skills, and stress levels to automatically generate the best response strategy. For example, for customers with multiple past complaints, it generates a faster solution or a special apology based on previous responses. In the medical field, it automatically generates appropriate treatment guides and precautions based on the patient's medical history and allergy information. In the financial field, it generates rapid self-verification steps and compensation guides based on past fraud history and personal verification information. In the retail sector, based on past purchase and return history, it automates return and exchange procedures and offers discounts for future purchases. In the communications sector, it combines group notifications with individual responses when a failure occurs, such as generating individual messages like "Current failure. Estimated recovery time is 0:00. We apologize for the inconvenience." Generative AI provides real-time prompts to operator terminals with generated response strategies and texts, additional proposals, precautions, FAQ links, and relevant manuals. For example, the operator screen might display specific instructions like, "This customer has made the same inquiry 3 times in the past. Since an apology and discount coupon were offered last time for delivery delays, we recommend adding an apology and a detailed explanation of the delivery status this time." Furthermore, the AI-generated response text can be sent directly via chat or automatically responded through speech synthesis. While referring to the AI's proposals, operators can fine-tune their responses based on their own judgment. Specifically, this generative AI processing unit executes an inference graph that connects multiple neural network models in a pipeline-like manner. The first model in this generative AI processing unit is a sequence labeling model that extracts named entities such as product names, dates, times, and amounts from input text. This generative AI processing unit, as the second stage of the model, takes the extracted named entities and customer history vectors as input and executes a multilayer perceptron to classify the categories of the response strategy. This generative AI processing unit, as the third stage of the model, uses the classified categories and sentiment labels as conditional cues and generates the final response text through an autoregressive large-scale language model. For example, this generative AI processing unit receives the text "The TV I bought yesterday has a broken screen" and customer history of a good customer as input data.The generative AI processing unit, as the first output, extracts the product TV and status screen fragmentation. The second output determines the appropriate category for a free replacement. The third output generates a text expressing sincere apologies and promising to arrange immediate delivery of a new TV. This generative AI processing unit applies an inappropriate language filtering model to the generated text, automatically eliminating the risk of compliance violations. This processing does not rely on a single model; by coordinating multiple models specialized for each task, it suppresses highly information-processing techniques that produce unfounded outputs. Through this structure, the system achieves the technical effect of guaranteeing the accuracy and security of the generated responses and protecting the corporate brand. This system is suitable for use cases in pharmaceutical companies with stringent quality standards, automatically generating accurate and legally safe initial responses to side effect reports in their drug consultation windows.

[0054] The terminal interface provides an operator-friendly interface for handling inquiries through various channels such as chat, voice calls, emails, and video calls. It also allows for the display of AI proposals, editing of response documents, reference to customer information, viewing of FAQs and manuals, and input of feedback. For example, operators can send AI-generated response documents with a single click in chat, or the AI ​​can display response solutions in real time during voice calls, or automatically pop up FAQs and manuals. Furthermore, it provides an interface for operators to instantly input feedback such as "This proposal is appropriate" or "This proposal is not suitable for the current situation" regarding AI proposals. Specifically, this terminal interface uses a two-way real-time communication protocol to establish a communication channel with the server. Implemented as a single-page application, the terminal interface asynchronously updates the document object model without screen transitions. During voice calls, the terminal interface captures the operator's and customer's voices as separate streams and sends them concurrently to the server's speech recognition engine. The terminal interface synchronizes the textual data returned from the server and the AI's response solutions with the timestamps of the speech, displaying them on a scrolling timeline on the screen. This terminal interface dynamically assigns background colors corresponding to the reliability scores to each sentence of the AI-generated response. When the operator corrects the response, the terminal interface packages the textual differences in the corrected area and the time spent on the correction as metadata. This packaged feedback data is then sent to the server in a background process. This processing is not merely a screen display program, but functions as a highly sophisticated sensor interface for collecting quantitative data on human-machine interaction. Through this architecture, the system achieves the technical effect of visually reducing the operator's cognitive load and preventing operational errors. This system is suitable for use cases in omnichannel support centers that handle multiple chat sessions simultaneously, supporting operator context switching.

[0055] The Feedback Accumulation Department collects and manages various evaluation metrics acquired after customer interactions, including customer satisfaction scores, survey results, operator feedback, AI proposal adoption rates, response speeds, complaint rates, and repeat consultation rates. For example, when a customer rating is "very satisfied," or when an operator reports that "the AI ​​proposal is very accurate," this information is provided to the Generative AI Processing Department for self-learning and model optimization. Conversely, in cases of low satisfaction, or if an operator points out that "the proposal is inappropriate," the generative AI initiates a self-correction function, automatically readjusting or relearning model parameters. Furthermore, the Feedback Accumulation Department also utilizes this information for business improvement and optimization of educational programs. Specifically, this Feedback Accumulation Department stores the collected evaluation metrics in a time-series database, constructing a multi-dimensional online analysis and processing cube. This department comprehensively analyzes quantitative data such as AI proposal adoption rates and response speeds, as well as qualitative data such as free-response questionnaires. The department uses a natural language processing model to extract themes and sentiment polarities from the free-response text, converting them into numerical sentiment scores. This feedback accumulation unit uses these numerical indicators as input features to construct a regression model that predicts customer satisfaction using a random forest algorithm. The unit calculates the importance of the features in the regression model, statistically identifying which response process contributes most to satisfaction. Based on these identified key factors, the unit executes a script that automatically rewrites the prompt templates for the generative AI processing unit. For example, if the unit determines that the presence or absence of an empathetic phrase is strongly correlated with satisfaction, it dynamically adds a constraint requiring empathy to be expressed at the beginning of the prompt. This process replaces the loop of data analysis and business improvement performed by human analysts with a self-regulating parameter adjustment loop based on machine learning algorithms. Through this structure, the system leverages its technical effectiveness to continuously optimize overall system performance in response to changes in the external environment and customer preferences without human intervention. This system is suitable for use cases where, within the support window of a smartphone application with a short product lifecycle and daily changing customer responses, the highest level of response quality must be maintained.

[0056] The Security Management Department encrypts confidential data such as customer information, reception history, and operator information to prevent unauthorized access and information leakage. For example, it uses quantum-resistant cryptography to protect communication paths, meticulously sets access permissions to restrict access to personal information, or records login and operation history as audit logs. Furthermore, it thoroughly protects privacy by setting a structure that sends only the minimum necessary information to the terminal, displaying only response policies and necessary historical information in a way that cannot be used to target specific individuals. Specifically, the Security Management Department implements an access control mechanism based on a zero-trust architecture. When operators log in, in addition to password authentication, the department requires multi-factor authentication using biometric authentication and hardware tokens. The department dynamically evaluates the operator's terminal's network address, physical address, and operating system patch availability to calculate a reliability score. If the reliability score does not meet the specified threshold, the department automatically revokes the operator's query execution privileges. The department applies preserved-format encryption to confidential fields within the database. The department also incorporates homomorphic cryptography that enables data retrieval and partial consistency comparison while maintaining encryption. When sending data to terminals, this security management department uses differential privacy algorithms to add intentional noise to the data, making it mathematically difficult to identify individuals. This process is an improvement in computer technology, replacing existing security models that rely on perimeter defenses with data-centric, dynamic, and cryptographic protection mechanisms. Through this architecture, the system achieves a technical effect that minimizes the risk of data breaches caused by insider crime or highly sophisticated cyberattacks. This system is applicable to use cases where it leverages AI, such as citizen consultation windows in government agencies handling highly confidential personal information or the anti-fraud investigation departments of credit card companies, while meeting stringent compliance requirements.

[0057] The External Liaison Department collaborates with external systems such as emotion recognition engines, speech recognition engines, CRM systems, FAQ databases, manual management systems, questionnaire systems, disaster information systems, and legal databases to acquire and utilize necessary information in real time. For example, in voice calls, the speech recognition engine transcribes the data, the emotion recognition engine determines the emotion, the department retrieves the latest customer information from the CRM system, extracts relevant information from the FAQ database, and obtains the latest response steps from the manual management system. During disasters, it retrieves information about affected areas from the disaster information system and the latest information from the legal database when laws and regulations change, automatically generating response documents. Specifically, this External Liaison Department functions as a gateway for asynchronous communication with multiple external systems using an application programming interface (API). It integrates response data from various external systems, converting it into standardized data objects that conform to a common pattern. It temporarily stores text blocks received streaming from the speech recognition engine in a memory buffer and pushes them to the generative AI processing department at regular intervals when sentence breaks are detected. It parses data in markup language format provided by the disaster information system, extracting parameters such as the epicenter and alarm level. This external linkage unit vectorizes the extracted parameters and performs geospatial distance calculations between them and the customer's residential coordinates. It identifies customers within a certain distance range and dynamically assigns disaster impact indicators to their sessions. This external linkage unit detects updates to the legal database via Webhook, automatically retrieves the updated clause text, and reconstructs the vector database index. This process is a highly integrated system technology that absorbs differences in data formats between isolated heterogeneous systems and builds real-time pipelined data. Through this structure, the system effectively adapts to rapid changes in the external environment without waiting for human data input, enabling the overall system behavior to change instantly. This system is suitable for use cases in customer centers of logistics companies operating globally, where it can obtain real-time weather information and customs regulation changes from various countries and automatically update delivery delay guidance.

[0058] As specific examples of prompts read by generative AI when implementing this invention, examples include: "Customer ID 12345 has inquired about delivery delays twice in the past, and an apology and discount coupon were provided last time. This time, the inquiry is about 'being troubled by the undelivered goods.' Based on past response history and customer attributes, the optimal apology and solution are generated." "Customer ID 67890 wants to change a medical appointment and has a history of diabetes. Based on appointment history and medical history, appropriate treatment guides and precautions are generated." "Operator A is skilled in the medical field and is currently under high stress. The inquiry is about 'explanation of test results.' Considering the operator's workload, a concise and accurate explanation is generated." "Inquiries about new services have increased dramatically. Referring to the FAQ database and past inquiry history, frequently asked questions and their answers are automatically generated." "A communication failure has occurred, requiring simultaneous notification of customers in disaster-stricken areas. The latest information is obtained from the disaster information system, and individual messages are generated." Specifically, this system inputs the prompts into a large-scale language model. This model tokenizes the input text. This model extracts feature vectors from the token sequence. This system generates the optimal response text based on the extracted feature vectors. This system outputs the generated response text. This processing improves the accuracy of response generation.

[0059] In this embodiment, the system comprises a customer information management department, an operator information management department, a generative AI processing department, a terminal interface department, a feedback accumulation department, a security management department, and an external linkage department. These departments work together to monitor the context and emotions of each customer in real time, automatically generating and suggesting optimal response strategies. This enables a highly efficient contact center system capable of continuous self-learning, self-correction, and environmental adaptation. Generative AI plays an indispensable role in all processes, including consultation analysis, emotion recognition, response generation, self-learning, self-correction, environmental adaptation, operator training, trend analysis, automatic FAQ generation, automatic manual updates, simultaneous notification during disasters, and response to changes in laws and regulations. It provides highly accurate and flexible customer responses that were previously impossible with rule-based systems. Specifically, this system employs an architecture that links multiple computing nodes in a distributed computing environment. The system containerizes each management department and processing department as independent microservices. The system uses container orchestration tools to dynamically allocate computing resources based on the load of each module. The generative AI processing department is built around a large-scale language model with a self-attention mechanism and hundreds of billions of parameters. This generative AI processing unit tokenizes the input natural language text using a sub-word algorithm. It embeds the tokenized sequences into a 1024-dimensional high-dimensional vector space. The unit then uses a multi-layered Transformer encoder and decoder to compute contextual dependencies. Unlike previous rule-based keyword matching, this system quantifies semantic similarity through inner product calculations in the high-dimensional vector space. This unconventional approach enables high-precision parsing even for queries involving word swings or complex contexts. This processing eliminates the mental process of manual information retrieval and comparison by human operators. As a result, the system significantly improves the speed of query response. To reduce communication overhead, the system also supports edge computing with lightweight inference models executed on the terminal side. This system demonstrates exceptional technical effectiveness in real-time applications such as fraud detection in the financial sector and triage in the medical field. During its self-learning process, the system continuously updates the model weights using a cross-entropy error function. This endows the computer system with adaptability to environmental changes. The system expands the updated parameters into memory and immediately reflects them in subsequent inference processing.

[0060] (Implementation steps) Step 1: Inquiry Acceptance and Customer Authentication. When a customer inquires through various channels such as chat, voice call, email, or video call, the terminal interface department receives the inquiry content and sends the customer ID and authentication information to the server. For example, in the financial field, a date of birth or registered phone number is required for identity verification; in the medical field, authentication is performed using a medical card number or date of birth. In the retail field, purchase history or membership number is used; and in the telecommunications field, contract number or phone number is used for customer authentication. Specifically, this terminal interface department uses an asynchronous communication protocol to receive input streams from multiple channels in parallel. This terminal interface department converts the received voice data into a digital signal in pulse code modulation format with a sampling rate of 16kHz. This terminal interface department encodes the text data as a byte sequence in UTF-8 format. This terminal interface department applies a preprocessing filter to remove noise components from the received data. This terminal interface department converts the preprocessed data into a 256-dimensional feature vector. This terminal interface department generates the extracted feature vector and an authentication token encrypted using a secure hash algorithm. This terminal interface department sends the authentication token to the server via the Secure Sockets Layer. This server decrypts the received authentication token. It then rapidly compares the token with the customer ID indexed on a distributed database. During this comparison, the server performs a retrieval using a B-tree algorithm. This allows the server to identify the customer from millions of customer records in milliseconds. Through this hardware-optimized data structure and retrieval algorithm, the system minimizes the latency of the authentication process. As a result, the system achieves a significant improvement in overall system response speed. This system is applicable to high-security use cases such as multi-factor authentication in the financial sector and biometric data verification in the medical field. After authentication, the system issues a trigger signal to transfer processing to the subsequent step 2. This trigger signal is registered in a message queue, and asynchronous processing continues.

[0061] Step 2: Extraction of Customer and Operator Information. The Customer Information Management Department on the server immediately extracts the consultation history, attribute information, past reception results, purchase history, complaint history, questionnaire results, usage channels, contract content, and past satisfaction scores of certified customers. Simultaneously, the Operator Information Management Department extracts the skill sets, responsibility history, stress levels, scope of business to be handled, operational status, rest status, past customer reviews, education history, and qualification information of all operators. For example, in the medical field, it extracts past medical history and appointment history; in the financial field, it extracts past transaction history and whether there has been fraudulent use; in the retail field, it extracts purchase and return history; and in the telecommunications field, it extracts contract packages and fault handling history. Specifically, this Customer Information Management Department utilizes a hybrid storage architecture combining relational databases and in-memory caching. This Customer Information Management Department receives the trigger signal issued in Step 1. Using the customer's ID as the key, this Customer Information Management Department extracts all relevant historical data from the distributed database through parallel queries. This Customer Information Management Department normalizes the extracted text data, numerical data, and categorical data respectively. The Customer Information Management Department combines the normalized data to generate a 1024-dimensional multidimensional tensor representing the customer profile. The Operator Information Management Department simultaneously acquires the real-time status information of all operators in operation. The Operator Information Management Department quantifies each operator's skill level as a continuous value from 0 to 1. The Operator Information Management Department obtains stress levels as a time-series array calculated from heart rate variability data or past response counts. The Operator Information Management Department integrates this information to generate a 512-dimensional state vector representing each operator's state. The system stores the customer profile tensor and operator state vectors in a shared memory space. This allows the subsequent generative AI processing unit to read data at high speed without disk I / O. Through this optimized memory access, the system significantly reduces the latency required to transition from data extraction to parsing. This system provides an indispensable foundational technology for use cases requiring instantaneous processing of massive amounts of data, such as real-time recommendation generation in the retail sector and rapid staffing during large-scale failures in the communications sector. The system serializes the extracted tensor data in conjunction with the input format of subsequent step 3.

[0062] Step 3: Consultation Analysis and Emotion Recognition Based on Generative AI. The generative AI processing unit parses the consultation content sent from the terminal interface unit using natural language and accurately identifies the emotions in the consultation text. For example, in the case of voice input, the text is transcribed by a speech recognition engine, and the emotion recognition engine identifies emotions such as "anger," "confusion," "gratitude," and "anxiety." Specific examples of prompts for the generative AI to read include: "Customer ID 12345 has inquired twice about delivery delays in the past, and an apology and discount coupon were provided last time. This time, the consultation content is 'I'm upset that the goods haven't been delivered.' Based on past response history and customer attributes, the best apology and solution are generated," and "Customer ID 67890 wants to change a medical appointment and has a history of diabetes. Based on appointment history and medical history, appropriate treatment guidance and precautions are generated," etc. Specifically, this generative AI processing unit drives a multimodal deep neural network to process the input consultation text and speech features. This generative AI processing unit converts text data into an array of integer tag IDs using a sub-word tagger. This generative AI processing unit converts speech data into a two-dimensional image tensor composed of Mel-frequency cepstral coefficients. It executes a Transformer encoder for text and a convolutional neural network for speech in parallel. The unit inputs the 768-dimensional hidden state vectors from the outputs of both networks into a concatenation layer. It applies a self-attention mechanism to the concatenated vector to generate a multimodal feature vector. This feature vector is then input into a fully connected layer for sentiment classification. The unit applies a Softmax function to output probability distribution scores from 0.0 to 1.0 for each sentiment category, such as "anger," "confusion," "gratitude," and "anxiety." For example, given the input text "I'm annoyed that the goods haven't been delivered" combined with a speech pitch indicating anxiety, the unit outputs a probability score of 0.85 for "confusion" and 0.12 for "anger." The unit then determines the sentiment category with the highest probability score as the final sentiment label. This system achieves objective and highly accurate emotion recognition without relying on human subjectivity through nonlinear operations in this high-dimensional space. This processing method enables the extraction of implicit emotions that traditional keyword matching methods could not detect. The system writes the output emotion tags and probability scores into memory as input parameters for subsequent optimal operator allocation and response generation processes. This system is indispensable for use cases requiring accurate emotion understanding, such as early complaint detection in the financial sector and triage based on patient anxiety levels in the medical field.

[0063] Step 4: Optimal Operator Allocation. The generative AI processing unit comprehensively assesses the consultation content and customer attributes, operator skill set and stress level, operational status, qualification information, etc., and automatically allocates the optimal operator. For example, for consultations requiring medical expertise, operators with medical skills are prioritized; for operators with high stress levels, automatic adjustments are made to reduce their workload. Specifically, this generative AI processing unit receives the sentiment labels and probability scores output in Step 3, and the customer profile tensor generated in Step 2 as input. This generative AI processing unit executes a reinforcement learning-based routing algorithm. This generative AI processing unit obtains the state vectors of each currently operating operator. This generative AI processing unit calculates the cosine similarity between the skills required by the customer and the skills possessed by the operator. This generative AI processing unit incorporates a numerical value representing the operator's current stress level and a fatigue score based on past coping history as penalty terms into the objective function. This generative AI processing unit calculates the ID of the operator that maximizes the objective function. For example, in consultations requiring medical expertise, this generative AI processing unit extracts operators with medical skills scores above 0.9 and current stress level scores below 0.4 as candidates. When multiple candidates exist, the unit compares the resolution times of similar past cases and ultimately selects the operator with the shortest resolution time. This system, through its multi-objective optimization algorithm, avoids random allocation based solely on idle conditions, simultaneously achieving both overall system processing efficiency and operator workload distribution. Utilizing the computer's unique computational processing capabilities, this system achieves the technical effects of reducing waiting time and homogenizing response quality. The system sends the ID and routing score of the best operator to the session management module. The session management module dynamically establishes communication paths based on this ID and transfers processing to subsequent step 5. This system is applied to priority-based routing during large-scale failures in the communications field and resource allocation under complex constraints in the retail sector, such as assigning dedicated personnel to VIP customers. The system logs the allocation results and utilizes them as relearning data for future routing models.

[0064] Step 5: Automatic Generation of Response Strategies and Response Texts Based on Generative AI. The generative AI processing unit comprehensively considers information such as the customer's past history and attributes, consultation content, operator skills, and stress levels to automatically generate optimal response strategies and response texts, supplementary proposals, precautions, FAQ links, and relevant manuals. For example, it generates specific instructions such as, "This customer has made the same inquiry 3 times in the past. Since an apology and discount coupon were provided for the delivery delay last time, we recommend adding an apology and a detailed explanation of the delivery status this time." Specific examples of prompts for the generative AI to read include: "Customer ID 24680 has made 2 returns in the retail sector in the past. This inquiry is 'I want to return the item.' Based on past return history, generate the optimal return procedure guide and precautions," and "In the event of a failure in the communications sector, customers in the affected areas need to be notified simultaneously. Obtain the latest information from the disaster information system and generate individual messages." Specifically, this generative AI processing unit drives an autoregressive language generation model. This generative AI processing unit combines a customer profile tensor, a labeled array of consultation texts, extracted sentiment tags, and assigned operator skill information to construct a context vector. This context vector is then embedded into a prompt template to generate input prompts for the model. In each layer of the Transformer decoder, this generative AI processing unit uses a masked self-attention mechanism to calculate the probability of the next label appearing from past labels. A beam search algorithm is applied to explore the label sequence with the highest probability product, generating a natural language response text. Simultaneously, this generative AI processing unit uses information retrieval-based generation techniques to retrieve vector representations of relevant documents from an external FAQ database. The content of the retrieved documents is integrated into the generation process to construct an accurate, fact-based response strategy. As output data, this generative AI processing unit generates the text string of the response text, structured JSON data representing recommended response steps, and a group of URLs for relevant manuals. For example, when handling inquiries about return procedures, this generative AI processing unit outputs structured data including "return application number," "return address," and "precautions," along with a text of an apology that resonates with the customer's emotions. This system, through a probabilistic generation process in a high-dimensional vector space, does not mimic human thought processes but synthesizes optimal text using a non-traditional method based on a statistical language model. This eliminates the time operators spend creating text from scratch, significantly reducing response generation latency. For subsequent display processing in step 6, the system forwards the generated output data to the terminal interface unit. This system is applicable to use cases requiring accuracy and personalization, such as the automatic generation of complex contract descriptions in the financial sector and the creation of personalized medication instructions in the medical field.

[0065] Step 6: Proposal and Response Implementation to Operator Terminals. The response strategies and responses generated by the generative AI processing unit, along with additional proposals, precautions, FAQ links, and relevant manuals, are provided to the operator terminal in real time. While referring to the AI's proposals, the operator can fine-tune the response content according to their own judgment and respond to the customer via chat, voice, email, etc. For example, the AI-generated response can be sent with one click, or an automatic response can be provided through speech synthesis. Specifically, the terminal interface unit asynchronously receives the response text, structured data, and URL link group generated in Step 5 from the server. The terminal interface unit renders the received data into a designated widget on the graphical user interface. The terminal interface unit configures the response text in an editable text box, allowing for real-time corrections by the operator. When the terminal interface unit detects that the operator has pressed the one-click send button, it datagrams the string in the text box and sends it to the client terminal. If the speech synthesis function is enabled, the terminal interface unit activates the text-to-speech engine. The terminal interface unit converts the response text into a phoneme array and generates a Mel-spectrum using a deep acoustic model. The terminal interface uses a vocoder to convert the Mel spectrogram into time-domain speech waveform data, which is then played back via an audio output device. In this series of processes, the system performs parallel computations using a graphics processing unit, suppressing the latency of image rendering and speech synthesis to the millisecond level. Through this close integration of hardware and software, the system reduces the physical workload of operators and prevents human error. When the operator corrects the text, the system extracts the differential data and records it in a log file. This log data is used as feedback information in subsequent step 7. This system is applied to use cases integrating multiple communication channels, such as automated voice responses for standardized fault guidance in the communications field and seamless interaction between chatbots and humans in the retail sector. After sending, the system updates the session state and transitions to the next consultation standby state.

[0066] Step 7: Feedback Collection and Accumulation. After customer reception, various evaluation indicators, such as customer satisfaction score, questionnaire results, operator feedback, AI proposal adoption rate, response speed, complaint rate, and repeat consultation rate, are sent from the terminal interface department to the feedback accumulation department for accumulation and management. For example, if the customer rating is "very satisfied," or if the operator feedback is "the AI ​​proposal is very accurate," this information is recorded. Specifically, when the terminal interface department detects the end of a customer reception session, it initiates the evaluation data collection module. The terminal interface department sends a score input form with five levels of evaluation to the customer terminal, receiving the input values ​​as the customer satisfaction score. On the operator terminal, the terminal interface department obtains the binary value indicating whether the AI-generated response text was adopted, as well as the string representation of the corrected text. The terminal interface department measures the elapsed time from the start to the end of the session as a numerical data in milliseconds and records it as the response speed. The terminal interface department formats this collected data into a structured JSON object associated with the session ID, customer ID, and operator ID. The terminal interface department sends this JSON object to the feedback accumulation department via a secure communication channel. The feedback accumulation unit writes the received data into a time-series database. During the write process, transaction control is implemented to ensure data consistency. This system, through its automated data collection mechanism, eliminates data gaps or inconsistencies caused by manual human input. By persisting large amounts of evaluation data in a structured format, this system effectively streamlines the relearning process of subsequent machine learning models. The accumulated data is provided as the input dataset for the self-learning algorithm in step 8. This system is applied to use cases requiring evidence preservation and quality management, such as post-compliance audits in the financial sector and continuous monitoring of hospitality quality in the medical field. The system automatically triggers the learning pipeline when the accumulated data exceeds a certain threshold.

[0067] Step 8: Self-learning, self-correction, and environmental adaptation based on generative AI The generative AI processing unit analyzes evaluation metrics, new consultation patterns, and changes in the market environment obtained from the feedback accumulation unit, continuously performing self-learning to optimize model parameters and response generation algorithms. For example, if satisfaction is low in a specific response pattern, it automatically corrects and improves the response text and steps for that pattern, and also promotes the high-satisfaction pattern to other operators. Specific examples of prompts for the generative AI to read include: "Analyze customer satisfaction scores and AI proposal adoption rates over the past month to correct response texts for patterns with low satisfaction," and "Inquiries about new services are increasing rapidly. Refer to the FAQ database and past consultation history to automatically generate common questions and their answers." Specifically, this generative AI processing unit uses scheduled batch processing or a specific amount of accumulated data as triggers to start the self-learning pipeline. This generative AI processing unit extracts a group of structured JSON objects from the feedback accumulation unit over a certain period. From the extracted data, this generative AI processing unit extracts data on sessions where the AI ​​proposal adoption rate is below a threshold (e.g., 0.5) or sessions with low customer satisfaction scores. This generative AI processing unit constructs pairs between input hints from these failure cases and the correct answer text ultimately corrected and sent by the operator. Using these pairs, the generative AI processing unit performs fine-tuning of a large-scale language model. During the learning process, the generative AI processing unit calculates the cross-entropy loss between the generated text and the correct answer text. The generative AI processing unit uses backpropagation to incrementally update the network's weight parameters. To prevent overfitting, the generative AI processing unit applies regularization techniques such as gradient clipping and Dropout. After learning is complete, the generative AI processing unit deploys the updated weight parameters as a new inference model in memory. This closed-loop machine learning process enables a non-traditional processing method where the system autonomously improves performance without requiring manual rule rewriting by human programmers. Through this self-correcting function, the system dynamically adapts to unknown consultation patterns and changes in the market environment, resulting in a technical effect that prevents long-term degradation of response accuracy. This system plays a crucial role in use cases with drastically changing environments, such as handling unknown problems during new product launches in the retail sector and learning new failure modes in the communications field. This system uses the updated model to perform the inference process after step 3.

[0068] (specific use case) For example, in the telecommunications field, when a customer inquires about internet connection problems via chat, the terminal interface department receives the inquiry and sends the customer ID and contract number to the server. The customer information management department immediately retrieves the customer's past inquiry history, contract package, terminal used, past satisfaction scores, location of the problem, and past complaint history. The operator information management department retrieves all operators' skill sets, troubleshooting experience, stress levels, operational status, past customer reviews, and nighttime handling capabilities. The generative AI processing department analyzes the received inquiry content using natural language, judging the customer's emotions as "anger" or "confusion" from expressions such as "network connection lost" and "please don't make me repeat myself." The generative AI comprehensively considers the customer's past inquiry history and attributes, the current problem situation, and the operator's skills and stress levels to automatically generate the best response strategy and response text. For example, it might generate specific instructions and responses such as, "Currently, a problem has occurred in area 00. The estimated recovery time is 0:00. We apologize for the inconvenience. Since the last problem took 0 hours to recover, we recommend that you periodically check the progress." On the operator's terminal, real-time prompts are displayed, including AI-generated response texts, descriptions of the fault situation, FAQ links, progress guide templates, and apology examples. Operators refer to the AI's proposals and fine-tune their response texts as needed, guiding customers through chat. After customer interaction, customer satisfaction scores and operator feedback are sent to the feedback accumulation department. The generative AI uses this information to continuously learn and optimize fault response patterns and response texts. Specifically, this system has an external linkage unit that integrates with the communication network's monitoring system. The terminal interface unit receives chat messages from customers as asynchronous text streams. The customer information management unit queries a distributed database using the customer ID as the key, generating a multi-dimensional customer profile tensor containing past fault consultation history. The operator information management unit generates a state vector that quantifies each operator's expertise score in the communication domain and the current number of response sessions. The generative AI processing unit inputs the received text stream into a natural language processing model, generating word embedding vector representations. This generative AI processing unit uses the self-attention mechanism of the Transformer architecture to parse the context, outputting a probability score of 0.92 for the "anger" category and 0.85 for the "confusion" category from the sentiment classification layer. This generative AI processing unit compares the polygonal data of the current fault location obtained from the external linkage unit with the coordinate data of the customer's registered address through spatial operations to determine that the customer is affected by the fault. This generative AI processing unit integrates these input data sets to construct a context vector and inputs it into an autoregressive language model.This generative AI processing unit outputs a text string containing an apology, estimated recovery time, and a personalized progress guide based on past response history. In this series of processes, the system integrates information from multiple data sources in a high-dimensional space, performing flexible text synthesis impossible with rule-based branching processes. This eliminates the time operators spend navigating multiple system screens to collect information, significantly reducing average processing time. The system renders the generated text string and related FAQ URLs on the operator's terminal screen in real time. The system uses the evaluation scores collected after the reception to weight the loss function and updates the parameters of the next response generation model using a reinforcement learning algorithm. This mechanism allows the system to be applied to use cases that prevent call center congestion during large-scale communication failures and minimize customer dissatisfaction.

[0069] As specific examples of prompts that can be read by generative AI, consider the following: "Customer ID 12345 has made 3 previous fault inquiries in the communications field. This time, the inquiry is about 'network connectivity issues.' Currently, a fault has occurred in area 00, with an estimated recovery time of 0:00. Based on past fault handling history and customer attributes, generate the best apology and progress guidance text." Another example is, "Operator A has extensive experience in handling faults in the communications field, but recently has faced a high volume of cases and a high level of stress. This inquiry concerns the occurrence of a fault. Considering the operator's workload, generate a concise and accurate explanation text." Specifically, this generative AI processing unit executes a dynamic prompting engineering module. This module maps structured data extracted from the database to a prompt template in text form interpretable by a natural language model. This module retrieves the number of past customer inquiries as integer values, the name of the fault location as a string, and converts the estimated recovery time from a timestamp format into a readable string. This module inserts these variables into placeholders within the template to construct the input prompt string. This generative AI processing unit then uses a tokenizer to convert the constructed input prompt string into a multi-dimensional array of labeled IDs. This generative AI processing unit feeds the array of tag IDs into the input layer of a large-scale language model. Under high operator stress, this unit assigns metadata to the prompts, representing the constraint of "generating concise and accurate descriptive text." Based on this metadata, the unit applies a loss function that penalizes the length of the output tags in the decoder layer. This suppresses verbose expressions, outputting only short text strings summarizing the key points. This system employs an unconventional processing method—dynamic prompt generation and constrained decoding—to force the computer to output the optimal form of information appropriate to the situation. This process reduces operator cognitive load and prevents information omissions and communication errors. The system then passes the output text string to subsequent display rendering. This system is highly effective not only in fault handling in the communications field but also in high-pressure use cases requiring rapid and accurate information delivery, such as responding to sudden market changes in the financial sector. The system saves the history of generated prompts and utilizes this data for future template optimization analysis.

[0070] In the medical field, when a patient wishes to change their appointment, the terminal interface department sends the consultation content and patient card number to the server. The customer information management department extracts appointment history, past medical history, allergy information, past treatment content, satisfaction scores, etc. Generative AI, referencing the patient's past medical history and appointment history, automatically generates appropriate treatment guidance and precautions using a prompt such as, "A patient with a history of diabetes wishes to change their appointment. The current appointment is on [Date] at [Time]. Please check availability and guide the doctor to change if necessary." On the operator's terminal, the AI-generated appointment change guidance, precautions, doctor change proposals, and FAQ links are displayed. The operator, referring to the AI's proposals, provides optimal guidance to the patient. Specifically, this system links with the medical information system via a secure API. The terminal interface department streams voice data containing the patient's appointment change request and converts it into a text string using a real-time speech recognition engine. The customer information management department accesses the electronic medical record database using the patient card number as the key, extracting structured medical data records containing past medical history, allergy information, and past prescription drug lists. This generative AI processing unit inputs extracted medical data records into a natural language processing model, mapping them to concept vectors using a medical terminology ontology dictionary. When the generative AI processing unit detects that a patient's medical history includes "diabetes," it collaborates with a rule-based inference engine to establish constraints regarding dietary restrictions and the timing of insulin administration. This unit integrates these constraints, the current appointment date and time, and calendar data representing the doctor's available time slots to generate a context vector. This context vector is then input into a large-scale language model, outputting a list of alternative appointment date and time candidates, a text string containing specific precautions for diabetic patients, and a proposed reasoning for changing the responsible doctor as needed. This system does not simply retrieve available time slots on the calendar; rather, it addresses a highly sophisticated scheduling problem that considers the patient's medical background within a multi-dimensional vector space. This results in a technical effect of reducing the risk of medical accidents and improving the efficiency of appointment change procedures while ensuring patient safety. The system displays the generated text string and candidate list in a structured manner on the operator's graphical user interface. This system performs the transaction processing of writing the final appointment date and time selected by the operator back to the electronic medical record database. This system is applied to use cases requiring medical expertise and complex adjustments, such as remote medical consultations and the scheduling of regular appointments for patients with chronic diseases.

[0071] Furthermore, in the retail sector, when a customer wishes to return an item, the terminal interface department sends the inquiry and member ID to the server. The customer information management department then extracts purchase history, past return history, complaint history, satisfaction scores, and channel usage. Generative AI uses a prompt message such as, "Customer ID 24680 has made two returns in the retail sector in the past. This inquiry is about 'wanting to return an item.' Based on past return history, generate the best return procedure guide and precautions," automatically generating a return procedure guide, precautions, a FAQ link, and previous responses. On the operator's terminal, the AI-generated return guide and precautions are displayed, allowing the operator to quickly and accurately guide the customer. Specifically, this system synchronizes in real-time with the e-commerce platform's order management database. The terminal interface department receives chat messages containing customer return requests. The customer information management department generates a multi-dimensional customer profile tensor using the member ID as the key, containing past purchase history, return history, and product category information. This generative AI processing unit parses the input chat messages using a natural language processing model, outputting probability distribution scores for classifying return reasons (e.g., "size discrepancy," "initial defect," etc.). It then detects past returns exceeding a threshold (e.g., 2 times) from the customer profile tensor. This detection result and the classified return reasons are integrated into a context vector and input into an autoregressive language model. As output, the unit generates text strings outlining specific return steps based on the return policy, structured data for the return address, and notes based on past return tendencies (e.g., "Do not cut off tags," etc.). In this process, the system analyzes past customer behavior patterns in the vector space, performing a non-traditional method of dynamically synthesizing and optimizing guide messages for individual situations. This eliminates the time operators spend referring to manuals to confirm return conditions, minimizing customer waiting time. The generated guide messages and structured data are displayed on the operator's terminal screen. After the operator sends the guide message, the system issues an API call to automatically update the return processing status to the order management database. This system achieves extremely high business efficiency in retail use cases that require frequent size changes in the apparel industry and troubleshooting initial defects in home appliance stores, where a large volume of transactions and rapid responses are essential. The system also periodically exports statistical data on return reasons to the marketing department's dashboard.

[0072] Thus, the system of this invention comprehensively utilizes each customer's history and attributes, consultation content, operator skills, and operational status across diverse industries such as communications, healthcare, retail, and finance. It uses generative AI to automatically generate optimal response strategies and texts, providing real-time prompts on operator terminals, thereby improving customer satisfaction, increasing business efficiency, and reducing operator workload. Specific examples of prompts read by the generative AI include: "Customer ID 13579 has previously inquired about fraud in the financial sector. This inquiry concerns 'card fraud.' Based on past response history and customer confirmation information, generate the optimal customer confirmation steps and compensation guidance text," and "Inquiries about new services are increasing rapidly. Based on the FAQ database and past consultation history, automatically generate frequently asked questions and their answers." Specifically, this system employs an architecture that parallelly assembles multiple specialized large-scale language models with domain knowledge specific to each industry. The system uses clustering algorithms to parse the vector representation of the input consultation text, automatically classifying it to belong to which domain (communications, healthcare, retail, or finance). Based on the classification results, the system routes processing to the corresponding specialized model. This generative AI processing unit, using a financial domain model, receives the customer's past transaction history tensor and current inquiry text as input. It extracts the word embedding vector for the keyword "fraudulent exploitation" from the text and initiates anomaly detection algorithms. As output, it generates structured data representing the steps to stop credit card use, a list of questions for self-verification, and a text string of compensation guidance to alleviate customer anxiety. The system also parses inquiry log data from a past period using a topic modeling algorithm in batch processing. When the system detects a sharp increase in the frequency of specific keywords related to new services, it automatically generates question-and-answer pairs related to that topic and updates the FAQ database index. Through its dynamic routing and autonomous knowledge base update mechanism supporting multiple domains, this system achieves a non-traditional approach to simultaneously meet the complex requirements of various industries with a single system. This results in a combined technical effect of reducing initial system implementation costs while improving customer satisfaction and reducing operator workload across industries. The system caches generated prompts and responses in memory, further reducing latency in subsequent inference processing for similar inquiries. This system is applied to large-scale use cases that require scalability and versatility, such as global customer support for multinational corporations and unified contact centers for group companies with multiple business operations.

[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) This invention provides a nursing support system comprised of a server and multiple terminals, including a user information management department, an employee information management department, a generative AI processing department, a family interaction department, a feedback accumulation department, a safety management department, an external interaction department, and a terminal interface department. The server handles the core functions of these departments. The terminals provide user interfaces for nursing staff, users, and family members, and include, as needed, lightweight AI inference modules, voice input devices, vital sign sensor linkage functions, and notification functions. Specifically, this nursing support system employs a distributed processing architecture built on a cloud computing environment. The server containerizes each management and processing department as an independent microservice and executes it on a container orchestration platform. This allows the server to dynamically scale its computing resources horizontally based on changes in access load. The terminals include tablets configured within the nursing facility, smartphones carried by staff, and IoT devices placed at the user's bedside. The terminal interface includes a directional microphone array as a voice input device and performs digital signal processing to reduce ambient noise. This terminal, acting as a vital signs sensor linkage function, uses a low-power short-range wireless communication protocol to collect real-time time-series data such as heart rate, blood pressure, and body temperature from wearable devices. The collected data is input into a local lightweight AI inference module, using edge computing technology to perform one-time anomaly detection processing. Through this edge and cloud collaborative processing, the system achieves the technical effect of saving communication bandwidth to the server while reducing response latency for high-urgency events to milliseconds. Upon detecting an anomaly, the system uses a notification function to display a visual alarm on the employee's terminal screen, while simultaneously driving a vibration motor to provide tactile feedback. This hardware-level linkage replaces human surveillance with continuous computer-based monitoring, providing an unconventional approach to prevent oversights. This system provides a crucial foundation for use cases requiring limited personnel to ensure security, such as nighttime care in specialized nursing homes and remote monitoring in home care. All collected data is encrypted and stored in a secure storage area.

[0075] The User Information Management Department, implemented on a server, stores and manages each user's nursing history (e.g., bathing assistance, feeding assistance, medication management, rehabilitation implementation records, excretion assistance, night care, fall history, emergency response history, pressure ulcer care, oral care, outing support, interest activity support, dementia care records), medical information (e.g., past medical history, medication content, allergy information, attending physician information, vital signs data, examination results, vaccination history, medical institution visit history, medical treatment records), lifestyle habits and preferences (e.g., dietary preferences, wake-up and bedtime, interests, family composition, religious considerations, bathing frequency, drinking and smoking habits, outing frequency), emotional and physical condition change history (e.g., daily speech and facial expressions, sleep status, activity level, pain and anxiety requests, mood changes, smile frequency, rejection response, loneliness, decreased willingness), and family contact history (e.g., requests from family members, contact frequency, past report content, family meeting minutes, emergency contact information, family member's nursing involvement history), etc. For example, if a user reports "headache today," the system immediately extracts multiple relevant data points, including past headache complaints and medication history, vital signs, recent sleep patterns, food intake, stress factors, and changes in air pressure, and utilizes this information in AI-based appropriate response decisions. Similarly, if a user reports "lack of appetite recently," the system can recommend contacting a nutritionist or doctor based on past food intake and weight changes, medication history, bowel movement records, blood sugar levels, relevant medical information, and dietary requests from family members. Furthermore, if a user reports "insomnia at night," the system can automatically generate suggestions such as "consult a doctor," "report sleep status to family," and "instruct staff to strengthen nighttime care" based on past sleep records, medication history, daytime activity levels, caffeine intake, nighttime urination frequency, noise levels, lighting conditions, and past responses to insomnia complaints. Specifically, the user information management department uses a graph database to persistently store multi-layered user information as a network structure with nodes and edges. This User Information Management Department defines data points such as nursing history, medical information, lifestyle habits, and emotional history as nodes, and links them using temporal and causal relationships as edges. When receiving user voice data from the terminal, this department uses a speech recognition engine to convert it into a text string. This text string is then input into a natural language processing model to extract semantic feature vectors. The extracted feature vectors are used as query inputs to a graph database, where a exploratory algorithm rapidly extracts relevant past node clusters. For example, for the text input "headache," the department extracts text nodes related to past headache complaints, recent blood pressure data nodes, and time-series data nodes showing changes in air pressure.This generative AI processing unit integrates the extracted node cluster data to construct a multidimensional context tensor. This context tensor is then input into a deep neural network, outputting a probability score representing the recommendation level of each action. The unit selects the action with the highest probability score (e.g., "recommend consulting a doctor") and generates structured JSON data containing its execution instructions. This system, through data extraction using graph structures and reasoning in a high-dimensional vector space, eliminates the mental processes of human caregivers relying on memory and paper records for judgment, enabling computer-based decision-making based on objective and comprehensive data. This results in the technical benefits of preventing the provision of incorrect care and reducing health risks to users. The generated JSON data is sent to employee terminals for subsequent notification processing. This system functions extremely effectively in use cases where complex factors intertwine in nursing settings, such as predicting wandering in dementia patients and early detection of signs of worsening chronic diseases.

[0076] The Employee Information Management Department manages each nursing staff member's skill set (e.g., bathing assistance, medication management, rehabilitation guidance, dementia care, medical care, emergency response, pressure ulcer care, oral care, excretion care, recreation planning, family management, infectious disease control), responsibility history (e.g., which users were cared for during which period, past response records, emergency response experience, family meeting participation history), stress level (e.g., recent workload, number of night shifts, feedback content, AI-based sentiment analysis, rest attainment, workload index, heart rate variability data), and scheduling information (e.g., ...). Information includes: day shift, night shift, rest days, emergency shift change history, expected rest days, overtime history, qualification information (e.g., nursing welfare worker, nurse, dementia nursing specialist, first aid and lifesaving training completion, infectious disease countermeasures training completion, recreational instructor), past evaluations (e.g., evaluations from users and families, evaluations from managers, AI-based reception quality evaluation), workload (e.g., current number of users under management, number of emergency cases being handled, workload for each business category, number of emergency response cases), and education history (e.g., training courses attended, OJT records, e-learning history, qualification acquisition progress), etc. For example, real-time monitoring of information such as employee A's expertise in dementia nursing, employee B's ability to provide medical care, employee C's new employee with high stress levels, employee D's night shift specialist status, and employee E's expertise in infectious disease countermeasures can be used for work allocation and workload adjustment, education plan development, and optimal staffing in emergencies. For example, in emergencies, employees capable of providing medical care can be prioritized. For example, when employees experience high stress levels, the system automatically sends alerts to management terminals such as "Recommended Rest," "Workload Reduction Plan," and "Recommended Mental Health Counseling." Specifically, the Employee Information Management Department uses a relational database with a standardized table structure to manage the attribute data of each employee. This department represents each employee's skill set as a multi-dimensional binary vector and records their workload and workload as a time-series numerical array. The department continuously receives heart rate variability data from wearable sensors on employee terminals and uses a frequency analysis algorithm to calculate stress index values ​​representing the balance between the sympathetic and parasympathetic nervous systems. This department integrates this data and updates the state vector representing each employee's current state in real time. In emergency situations, the generative AI processing department calculates the cosine similarity between the required skill vector and the state vectors of all employees. The generative AI processing department extracts employees with high similarity scores and current workload values ​​below a threshold as the best candidates. The generative AI processing department also calculates the moving average of each employee's stress index values ​​in periodic batch processing. If the moving average exceeds the specified safety threshold, this generative AI processing unit determines it to be in an abnormal state, generates a text string containing a warning message and structured data containing a recommended scheduling change plan, and sends it to the administrator's terminal.This system eliminates existing methods that rely on subjective judgment by managers through a dynamic resource allocation and monitoring mechanism based on quantitative data. This allows the system to effectively prevent employee overwork, reduce turnover, and maintain a certain level of overall facility care quality. The system caches the calculated state vectors in memory, minimizing latency in subsequent task allocation and processing. This system is applicable to use cases requiring dynamic resource management, such as complex scheduling in large-scale nursing facilities and emergency personnel allocation changes in zoning during infectious disease outbreaks.

[0077] The generative AI processing unit implements generative AI as an essential element on the server. Generative AI combines multiple AI technologies, including natural language processing models, large-scale language models, emotion recognition AI, predictive modeling AI, clustering AI, vital sign data analysis AI, image analysis AI, and speech recognition AI. It analyzes data from multiple perspectives, including user speech and log records, vital sign data, requests from family members, image data (e.g., photos of bedsores, photos of food), voice data (e.g., speech content, coughs, and breath sounds), and sensor information (e.g., fall sensors, bed exit sensors, door open / close sensors), to accurately identify changes in emotions and physical condition, abnormal signs, disruptions to daily routines, environmental changes, and changes in family members' requests. For example, if the user records a statement like "My chest hurts a little today," it is identified as "pain." The system then comprehensively considers past medical history and medication information, vital signs data, past responses to chest pain, recent activity levels, diet, stress factors, and changes in temperature and air pressure to automatically generate multiple nursing actions, such as "Contact medical staff to recommend vital sign checks," "Automatically send a status report to family members," and "Propose quiet instructions and enhanced observation to staff." Similarly, if the user says "Thank you, you've been a great help," it is identified as "gratitude," and a positive report such as "Today, the user expressed gratitude to staff" is automatically sent to family members, which is also used as data for improving staff morale and performance evaluation. Furthermore, if the user records statements like "anxiety" or "fear," the system automatically generates a proposal to "Contact psychological care staff and offer reassuring conversation," displaying specific instructions on the employee's terminal such as "Because the user is feeling uneasy, please speak gently and contact psychological care staff." For example, when image data of pressure ulcers is sent, the image analysis AI automatically determines the progression of the pressure ulcers and automatically generates "recommendations to doctors," "dressing change instructions," and "reports to family members." Similarly, when the fall sensor is activated, the AI ​​analyzes recent movement history and environmental factors, and immediately executes "fall cause analysis," "recurrence prevention strategy proposals," and "automatic reports to family members, doctors, and administrators." Specifically, this generative AI processing unit employs a multimodal Transformer architecture that integrates text, speech, images, and sensor time-series data. This generative AI processing unit converts the input speech waveform data into a spectrogram image tensor using Fourier transform and extracts acoustic feature vectors using a convolutional neural network. Simultaneously, this generative AI processing unit inputs image data such as pressure ulcers into a deep image recognition model, outputting a continuous score from 0 to 1 representing the progression of the lesion, as well as a feature map tensor. This generative AI processing unit receives binary signals from the bed exit sensor and fall sensor as a time-series event stream and uses a recurrent neural network to calculate the hidden state vector of the movement pattern.This generative AI processing unit uses a cross-attention mechanism to fuse feature vectors obtained from these different modalities into a single high-dimensional context vector. This context vector is then input into the decoder of an autoregressive language model, generating in parallel specific nursing instructions, reports for family members, and professional summary texts for medical staff. For example, when receiving a fall sensor's activation signal and recent anxious statements as input, the generative AI processing unit outputs a fall probability score of 0.98, instantly generating a command data packet instructing staff to proceed and sending it over the network. Through this multimodal data fusion and nonlinear computation in high-dimensional space, the system achieves a high-precision identification of complex anomalies that cannot be detected by a single sensor, employing unconventional processing methods. This system thus complements the limits of human patrol monitoring and prevents the escalation of accidents. The system forwards the generated text and command data packets to subsequent notification modules and the family coordination department. This system plays a crucial role in use cases requiring high levels of situational awareness and immediate response in nursing and medical collaboration, such as automating nighttime care and providing image-based diagnostic support from remote specialists.

[0078] The Family Liaison Department manages each family member's contact information, desired notification methods (e.g., email, app notifications, phone calls, chat), past contact history, requests, feedback from family members, family meeting minutes, family member participation history in care, family member health status, and care burden from multiple perspectives. It automatically sends AI-generated reports, emergency contact information, daily care summaries, user status changes, suggestions for family members, care burden reduction strategies, and educational content for family members to the families. For example, in the event of a user's physical condition changes or an emergency such as a fall, an immediate status report is sent to the family. It can also automatically generate weekly summaries of "this week's physical condition and emotional changes," "dietary intake trends," "sleep status," "recovery progress," and "words of gratitude to the family," and send them to the family. For example, if a family member requests to know "recent changes in physical condition," a summary of physical condition and emotional changes for the past week is automatically generated and sent to the family. Furthermore, when a family member's care burden is high, AI automatically proposes "recommendations for nursing leave," "guidance on regional support services," and "information on stress-related care for family members." Specifically, this family liaison department uses a message queue system to exchange data asynchronously with external communication networks. The department stores each family member's notification configuration file as a JSON document in a NoSQL database. In daily or weekly batch processing, the generative AI processing department statistically analyzes the user's vital sign data time series array, dietary intake data, and historical emotional tag data. This data is then input into a natural language generation model to generate a summary text string translated into colloquial language for family members. Simultaneously, the department uses a natural language processing model to parse past requests from family members, calculating weighted vectors for high-priority topics (e.g., "diet," "rehabilitation"). Based on these weighted vectors, the department dynamically optimizes the information configuration and emphasis within the summary text. The generated summary text and images of the user's smiles are packaged as a multi-part email payload or push notification payload. The department then sends the packaged payload via HTTPS to the API endpoint specified in the notification configuration file. This system fully automates the entire process from data collection to text generation and distribution, eliminating the time nursing staff spend manually filling out contact books. This significantly reduces the workload of staff while improving the frequency and quality of communication with families. Upon receiving a response from a family, the system parses the text data and adds it to the database as a new request, reflecting it in the next generation process.This system is applied to use cases that facilitate smooth information sharing among various stakeholders, such as regular status reports to family members living far away and instant escalation in emergencies.

[0079] The Feedback Accumulation Department collects and manages various evaluation indicators, including feedback from families and staff, users' physical condition and emotional change history, nursing implementation results, emergency response history, workload data, AI proposal adoption rate, response speed, complaint rate, re-response rate, family satisfaction, employee satisfaction, management evaluation, and external audit results. The Generative AI Processing Department utilizes this data for self-learning and model optimization, nursing procedure improvement, educational program optimization, business process improvement, risk management enhancement, and service quality improvement. For example, if families request that "recent reports are easy to understand" or "more detailed information is needed," the AI ​​automatically adjusts the report content and frequency. If staff provide feedback such as "AI proposals are not suitable for the situation" or "the proposals are accurate," the AI ​​automatically corrects the response generation algorithm to improve on-site suitability. For example, when the adoption rate of AI proposals is low, the AI ​​automatically performs cause analysis to improve the content, timing, and expression of the proposals. For example, when the re-response rate is high, the AI ​​analyzes past response history and automatically generates strategies to prevent relapse. Specifically, this Feedback Accumulation Department uses a distributed, large-scale data warehouse to centrally store both structured and unstructured data. The terminal interface department collects employee-input evaluation scores (e.g., integer values ​​from 1 to 5) for AI proposals and the revised text data of nursing records, sending them along with timestamps to the feedback accumulation department. The generative AI processing department uses the accumulated dataset to periodically execute reinforcement learning algorithms. The generative AI processing department calculates a positive reward signal when the nursing action proposed by the AI ​​is adopted and the user's vital signs stabilize afterward. The generative AI processing department calculates a negative reward signal when the proposal is rejected or a re-enforcement occurs. The generative AI processing department uses these reward signals to update the weight parameters of the neural network using the policy gradient method. For example, if the adoption rate of the proposal "provide hot drinks" for insomnia complaints during a specific time period is low, the generative AI processing department adjusts the output probability distribution for that context vector, increasing the probability score of other proposals such as "adjust ambient sound" in subsequent instances. Through this reward-based self-optimization mechanism, this system performs unconventional processing methods to automatically acquire tacit knowledge and adapt to the environment, achieving what static, rule-based systems cannot. This system achieves the technical benefit of continuously improving proposal accuracy and nursing quality over time. Updated model parameters are stored in a version control system and hot-deployed to the inference server. The system demonstrates exceptional effectiveness in value-generating use cases, such as initial optimization in newly established nursing facilities and learning patterns of user health changes with seasonal variations. The system visualizes summaries of the learning results and provides them to administrators on a dashboard.

[0080] The Security Management Department employs multi-layered encryption to protect confidential data such as user information, nursing history, medical information, family information, and employee information, preventing unauthorized access and information leakage. For example, it uses quantum-resistant cryptography to protect communication paths, finely sets access permissions to restrict access to personal information, or records login and operation history as audit logs. It can be configured to send only the minimum necessary information to the terminal, displaying only response strategies and necessary historical information in a way that cannot be used to target specific individuals. For example, usernames are anonymized and displayed on employee terminals, while no other user information is displayed on family terminals, ensuring thorough privacy protection. Specifically, this Security Management Department utilizes a dedicated encryption processor operating in a distributed computing environment. This encryption processor applies advanced encryption algorithms before converting input text and image data into high-dimensional numerical tensors. The system implements a quantum-resistant cryptographic algorithm based on lattice cryptography. The Security Management Department assigns dynamically generated 256-bit encryption keys to each data record. The system securely stores these encryption keys within a hardware security module. The Security Management Department receives access requests from terminals. The Security Management Department verifies the authentication token contained in the access request. This system only performs decryption processing if the authentication token is deemed legitimate. The security management department expands the decrypted data into a secure area in memory. Immediately after data processing, the system completely erases the data from memory. The security management department uses blockchain technology to record audit logs in a distributed ledger. This system makes tampering with audit logs mathematically impossible. The security management department inputs data sent to the terminal into an anonymization module. This anonymization module uses a natural language processing model to extract personally identifiable information from the text. This anonymization module replaces the extracted personally identifiable information with a virtual token. This system only sends anonymized text data via the network interface. The security management department also applies a face recognition model to image data. The face recognition model outputs the bounding box coordinates of the face region from the image tensor. This system applies Gaussian blur to this bounding box region. This system generates privacy-protected image data. These processes performed by the security management department result in a technical effect that minimizes the risk of information leakage. This system is extremely useful in use cases with stringent privacy requirements, such as medical institutions and nursing facilities. Unlike traditional simple password authentication and uniform encryption, this system implements dynamic and multi-layered security controls based on data content and access permissions. This system provides the highest level of security while optimizing computational resource consumption.

[0081] The External Liaison Department collaborates with multiple external systems, including medical databases, emotion recognition engines, speech recognition engines, AI-powered psychological care support systems, vital sign sensors, family contact systems, disaster information systems, regional integrated support centers, administrative services, long-term care insurance systems, and external auditing systems, to acquire and utilize necessary information in real time. For example, in the case of voice input, the speech recognition engine transcribes the data, the emotion recognition engine determines the emotion, the latest treatment guidelines and medication information are retrieved from medical databases, and real-time temperature, blood pressure, and pulse data are obtained from vital sign sensors, reflected in AI-based nursing proposals and risk assessments. For example, during disasters, information about affected areas is retrieved from disaster information systems, automatically generating safety confirmations and evacuation guides for users and their families. For example, in conjunction with external auditing systems, the department automatically evaluates compliance with regulations and service quality, providing improvement suggestions to managers. Specifically, this External Liaison Department has a gateway module that integrates and manages multiple application programming interfaces (APIs). This External Liaison Department uses asynchronous communication protocols to send and receive data packets with external systems. The system inputs received data in various formats into a standardization processing module. This normalization module converts JSON and XML data into a unified multidimensional tensor format. Upon receiving speech data, the external linkage unit converts it into a Mel-frequency cepstral coefficient feature vector. This feature vector is then input into a convolutional neural network for speech recognition. The neural network outputs a probability distribution of text tokens from the speech features. The system extracts the sequence of text tokens with the highest probability. The external linkage unit inputs the extracted text data into a Transformer model for emotion recognition. This Transformer model parses the text context using a self-attention mechanism. The Transformer model outputs confidence scores for emotion categories such as joy, anger, and sadness. The system identifies emotion categories whose confidence scores exceed a predetermined threshold. The external linkage unit receives time-series numerical array data from vital sign sensor devices. This numerical array data is input into a Long Short-Term Memory (LSTM) network. This LSM network learns the variation patterns of the time-series data. This LSM network outputs predicted scores and probabilities of future vital sign values. The system generates an alarm signal immediately when the probability of an anomaly exceeds a threshold. The external linkage unit dynamically generates and sends SQL queries to an external medical database. This system acquires the latest treatment guidelines and drug contraindications as structured data. This structured data is then integrated into a knowledge graph. The system uses the knowledge graph to perform reasoning processing. These processes, performed by the external linkage unit, result in the technical effect of real-time fusion of internal system data with the latest external information. This system is extremely useful in use cases such as rapid evacuation guidance during disasters and nursing proposals based on the latest medical insights.Unlike traditional batch-based data synchronization, this system uses stream processing technology to reflect external data into the system with millisecond-level latency. This system significantly improves the real-time performance and accuracy of data processing.

[0082] The Terminal Interface Department operates on various terminals (tablets, smartphones, voice input terminals, wearable terminals, PCs, and dedicated notification terminals) used by employees, users, and their families. It provides a variety of functions, including displaying AI proposals, editing response texts, inputting nursing records, acquiring vital sign data, sending image and voice data, confirming reports to family members, inputting feedback, viewing FAQs and manuals, listening to educational content, one-click notifications in emergencies, managing business progress, and alarm notifications. For example, it can send AI-generated response texts on chat screens with one click, display response plans in real-time during voice calls, or automatically pop up FAQs and manuals. For example, it can link with vital sign sensors to automatically perform AI analysis and issue alarms when abnormal values ​​are detected. For example, family members can use smartphones to check the user's latest condition and receive AI-generated weekly summaries and emergency reports instantly. Specifically, this Terminal Interface Department establishes bidirectional communication between the front-end application and the back-end server. This Terminal Interface Department uses the WebSocket protocol to maintain a continuous connection. The system will instantly distribute inference results and alarm information generated on the server side to the terminals as push notifications. This terminal interface unit possesses a rendering engine that dynamically optimizes the user interface layout based on the terminal's screen size and resolution. The terminal interface unit detects user touch operations and voice input via an event listener. The system inputs the voice data into a local lightweight acoustic model. This acoustic model converts the voice waveform into a text stream in real time. The system sends the converted text stream to a generative AI processing unit on the server side. The terminal interface unit displays candidate response texts received from the server on an overlay on the screen. The system uses the user's click event when selecting a response text as a trigger to determine that response text. The terminal interface unit acquires 3D motion vector data from the accelerometer and gyroscope sensors built into the wearable terminal. The system divides this motion vector data into a specific time window and generates a feature matrix. The system inputs this feature matrix into an inference module on the edge device. This inference module uses a support vector machine algorithm to output probability scores for abnormal actions such as falls. If the probability score exceeds a threshold, the system sends an emergency notification data packet to the server. These processes performed by the terminal interface unit reduce the user's operational burden and enable intuitive and rapid information input and output. This system is extremely useful in use cases such as hands-free operation at the nursing site and real-time status sharing with family members in remote locations. Unlike traditional interfaces based on static screen transitions, this system dynamically prompts the AI's reasoning results based on context. This system significantly improves the efficiency of human-computer interaction.

[0083] Thus, the embodiments of the present invention implement the core functions of a user information management department, an employee information management department, a generative AI processing department, a family interaction department, a feedback accumulation department, a safety management department, and an external interaction department on the server side, and configure employee, user, and family interfaces on the terminal side, as well as lightweight AI inference modules, vital sign sensor linkage functions, voice input functions, image sending functions, and notification functions as needed. In real time, from multiple perspectives, the system can grasp each user's physical condition and emotional changes, nursing history, family requests, employee status, environmental changes, and external risk information. It automatically generates and prompts the best nursing actions, reports, employee support, family interaction, risk management, educational support, and business improvement, realizing an advanced nursing support system that continuously performs self-learning, self-correction, environmental adaptation, and quality improvement. Generative AI plays an indispensable role in all processes, including consultation analysis, emotion recognition, response generation, self-learning, self-correction, environmental adaptation, family interaction, staff education, trend analysis, automatic generation of nursing procedures, simultaneous notification in emergencies, automatic updating of medical guidelines, risk assessment, external collaboration, and legal compliance support. This enables high-precision and flexible application in nursing settings and multifaceted value creation, which are impossible with traditional rule-based systems. Specifically, this system adopts a microservice architecture built on a cloud computing environment. Each management and processing department is deployed as an independent container. The system uses container orchestration tools to dynamically expand the computing resources of each module. The generative AI processing section utilizes a parallel computing cluster of graphics processing units. This system accelerates matrix multiplication operations in the inference processing of large-scale language models. As a lightweight AI inference module configured on the terminal side, this system compresses the model size of neural networks through quantization and pruning techniques. This system constructs a hybrid inference pipeline that combines primary inference at the edge and secondary inference at the cloud. This system only sends the feature vectors extracted at the edge to the cloud. This system significantly reduces network bandwidth consumption. It utilizes historical data stored in the feedback accumulation unit to execute reinforcement learning algorithms. The system defines user adoption and non-adoption history of nursing action proposals as reward signals. It continuously updates the parameters of the generative AI using policy gradient methods. The system autonomously improves the accuracy of proposals over time. These processes minimize overall system latency and ensure scalability and availability. This system is extremely useful in use cases such as large-scale nursing facility groups and wide-area regional integrated nursing networks. Unlike traditional static rule-based systems, this system dynamically adapts to environmental changes through a data-driven approach. It provides optimal solutions from a computer science perspective for the complexities of the nursing workplace.

[0084] (System Composition)The system involved in this implementation includes a user information management department, an employee information management department, a generative AI processing department, a family communication department, a feedback accumulation department, a security management department, an external communication department, and a terminal interface department. The user information management department stores and manages each user's nursing history, medical information, lifestyle habits and preferences, emotional and physical condition changes, and family contact history in a multi-layered and long-term manner. For example, it accumulates detailed nursing history records such as bathing care and dietary care, medication management, rehabilitation implementation records, excretion care, night care, fall history, emergency response history, pressure ulcer care, oral care, outing support, interest activity support, and dementia care records. It also comprehensively manages past medical history, medication content, allergy information, attending physician information, vital sign data, examination results, vaccination history, and medical institution visits. The system records medical information such as medical history and treatment records, as well as lifestyle habits and preferences such as dietary habits, wake-up and bedtime, interests, family composition, religious considerations, bathing frequency, drinking and smoking habits, and frequency of outings. It manages daily speech and facial expressions, sleep patterns, activity levels, complaints of pain and anxiety, mood changes, frequency of smiles, rejection responses, feelings of loneliness, and low motivation—all emotional and physical changes—in chronological order. It also comprehensively records family contact history, including requests from family members, frequency of contact, past reports, family meeting minutes, emergency contact information, and family participation in caregiving. For example, if a user reports "headache today," the system instantly extracts past complaints of headaches and medication history, vital signs data, recent sleep patterns, dietary intake, stress factors, and changes in air pressure, providing this data to the generative AI processing department. Specifically, this user information management department constructs a hybrid data storage system combining distributed relational databases and NoSQL databases. Vital signs and medication history, stored as structured data, are stored in tables within the relational database. This user information management department stores unstructured spoken text and freehand descriptions of nursing records in a document-oriented NoSQL database. The system assigns timestamps and user identifiers to each data record. The system uses a natural language processing model to extract named entities from the freehand text. The system vectorizes the extracted named entities and indexes them as coordinates in a high-dimensional vector space in a vector database. This user information management department receives the text data "Today I have a headache" from the terminal. The system converts this text data into a query vector. The system performs cosine similarity calculations on the vector database. The system quickly retrieves past nursing history vectors with high similarity to the query vector. The system combines the retrieved past historical data, a time-series array of vital sign data from the last 24 hours, and numerical data of air pressure obtained from environmental sensors. The system formats the combined data into a multi-dimensional context tensor. The system forwards this context tensor to the generative AI processing department.These processes performed by our User Information Management Department result in the efficient accumulation of vast amounts of data in various formats, and the ability to retrieve and extract necessary information in milliseconds. This system is extremely useful in use cases such as providing optimized personalized care for individual users and rapid information gathering in the event of sudden changes in physical condition. Unlike traditional manual paper records and simple electronic medical record retrieval, this system utilizes advanced information retrieval based on semantic similarity as internal computer processing. This system simultaneously achieves efficient data management and improved accuracy in information retrieval.

[0085] The Employee Information Management Department manages each nursing employee's skill set, responsibility history, stress level, shift schedule, qualifications, past evaluations, workload, and educational history from multiple perspectives. For example, it records detailed skill sets such as bathing care, medication management, rehabilitation guidance, dementia care, medical care, emergency response, pressure ulcer care, oral care, excretion care, recreation planning, family management, and infectious disease control. It also manages responsibility history, including which user the employee was responsible for for how long, past response records, emergency response experience, and family meeting attendance history. Furthermore, it tracks recent workload, number of night shifts, feedback content, AI-based sentiment analysis, rest attainment, and workload index. The system calculates stress levels using data such as heart rate variability, manages scheduling information including day shifts, night shifts, rest days, sudden scheduling changes, requested rest days, and overtime history, records qualification information for nursing welfare workers, nurses, dementia care specialists, completion of emergency care training, completion of infectious disease control training, and recreational instructors, accumulates past evaluations from users and families, managers, and AI-based reception quality assessments, manages current workload such as the number of users under management, the number of emergency cases being handled, the workload of each business category, and the number of emergency responses, and also records detailed educational history such as training sessions attended, OJT records, online learning history, and qualification progress. For example, it can monitor in real time whether employee A is skilled in dementia care, employee B can provide medical care, employee C is a new employee with high stress levels, employee D is a night shift specialist, and employee E has expertise in infectious disease control, and utilize this information for work allocation and workload adjustment, educational plan development, and optimal staffing in emergencies. Specifically, the Employee Information Management Department generates a multi-dimensional profile vector that integrates the attribute information and dynamic status information of each employee. The system converts skill sets and qualification information into binary vectors using one-hot encoding. Heart rate variability data and activity data obtained from wearable devices are input into the time series analysis module. This module uses Fast Fourier Transform to analyze the frequency components of heart rate variability. The system quantifies an indicator representing the balance of sympathetic and parasympathetic nervous system activity. This indicator, along with recent workload data, number of night shifts, and past feedback sentiment polarity scores, is input into a multiple regression analysis model. This model outputs a stress level score as a continuous value from 0 to 100. The calculated stress level score is dynamically reflected in the profile vector. The Employee Information Management Department monitors scheduling information and the current number of responsible users in real time. When a task assignment request occurs, the system executes an optimization algorithm. The system calculates the Euclidean distance between the request vector of the requested task and the profile vector of each employee. The system extracts the employee with the smallest distance and a stress level score below a predetermined threshold as the best candidate. The system will send the extracted employee identifiers and task details to the scheduling module.These processes performed by our Employee Information Management Department result in the technological benefits of quantitatively evaluating employee capabilities and well-being, and automating optimal resource allocation. This system is extremely useful in use cases such as complex scheduling in large-scale facilities and rapid staffing in emergencies. Unlike traditional staffing methods that rely on managerial experience and intuition, this system achieves scientific management based on objective data and mathematical optimization techniques. This system helps prevent employee turnover and maximize business efficiency.

[0086] The Generative AI Processing Unit implements generative AI as an essential element on the server, combining multiple AI technologies such as natural language processing models, large-scale language models, emotion recognition AI, predictive modeling AI, clustering AI, vital sign data analysis AI, image analysis AI, and speech recognition AI. The Generative AI Processing Unit analyzes data from multiple perspectives, including user statements and log records, vital sign data, requests from family members, image data, voice data, and sensor information. It accurately identifies changes in emotions and physical condition, abnormal signs, disruptions in daily routines, environmental changes, and changes in family requests. For example, if a statement like "My chest hurts a little today" is recorded, it is classified as "pain." The unit comprehensively considers past medical history and medication information, vital sign data, past responses to chest pain, recent activity levels, diet, stress factors, and changes in temperature and air pressure to automatically generate multiple nursing actions, such as "contacting medical staff to recommend vital sign checks," "automatically sending a status report to family members," and "suggesting quiet instructions and enhanced observation to staff." For example, when a user expresses "Thank you, you've been a great help," the AI ​​identifies it as "expressing gratitude" and automatically sends a positive report to the family member stating, "Today, the user expressed gratitude to the staff." This information is also used as data for improving employee morale and performance evaluation. For example, when a user records expressions of "anxiety" or "fear," the AI ​​automatically generates a suggestion to "contact the mental health staff and offer reassurance," displaying specific instructions on the employee's terminal such as, "Because the user is feeling anxious, please speak gently and contact the mental health staff." For example, when images of pressure ulcers are sent, the AI ​​automatically determines the progression of the ulcers and automatically generates instructions such as "Recommend reporting to the doctor," "Dressing change instructions," and "Report the progress to the family member." For example, when the fall sensor is activated, the AI ​​analyzes recent movement history and environmental factors, immediately executing actions such as "fall cause analysis," "recurrence prevention strategy proposals," and "automatic reporting to family members, doctors, and managers." As specific examples of prompts that generative AI can read, consider: "User ID U001 has a history of hypertension and is currently taking antihypertensive medication. Today's message is 'My chest hurts a little today.' Based on past nursing history and medical information, generate necessary nursing actions and a report to family members." "User ID U002 recently expressed 'anxiety.' Based on past emotional history and nursing records, generate examples of contacting psychological care staff and reassuring conversations." "Employee ID S001 has a high stress level. Based on past work records and feedback, generate an alert to management and a workload reduction plan." "Analyze pressure ulcer image data to generate a progress assessment and response proposal." "Analyze the action history and environmental factors when the fall sensor is working to generate a fall cause analysis and recurrence prevention measures."Specifically, this generative AI processing unit employs an ensemble learning architecture that integrates multiple neural network models. The system converts input text data into a permutation of sub-word tags using a word segmenter. This permutation is then input into a large-scale language model with billions of parameters. This large-scale language model uses the self-attention mechanism of the Transformer architecture to calculate long-distance dependencies between tags. Upon receiving image data, the system resizes and normalizes the image, converting it into a 3D image tensor. This image tensor is then input into a convolutional neural network. This convolutional neural network extracts feature maps from the image through multiple convolutional and pooling layers. Using the image tensor of pressure ulcers as input, the system uses a segmentation model to output pixel-level mask data of the affected area. This mask data and feature maps are input into a fully connected layer, which outputs a probability distribution representing the stage classification of pressure ulcer progression. The system determines the stage label with the highest probability. The determined stage label, the time-series prediction score of vital signs data, and the sentiment label obtained from text parsing are then input into the integration module. This integrated module combines this multimodal data and embeds it into placeholders in the prompt template. The system then re-inputs the constructed prompt string into a large-scale language model. This large-scale language model autoregressively predicts the next token, outputting specific nursing action instructions and family-oriented reports as text streams. The system then inputs the output text stream into a syntax parser, converting it into structured data in an executable command format. This structured data is forwarded to subsequent notification and database update modules. These processes performed by this generative AI processing unit result in a unified interpretation of diverse input modalities and the autonomous generation of optimal action plans appropriate to the situation. This system is extremely useful in use cases such as judgment support in healthcare and reducing employee workload. Unlike existing single rule-based decision-making systems, this system performs sophisticated reasoning that combines complex pattern recognition and contextual understanding in a high-dimensional space. While mimicking human cognitive processes, this system processes data with speed and comprehensiveness that surpasses them.

[0087] The Family Collaboration Department manages each family member's contact information, desired notification methods, past contact history, requests, feedback, meeting minutes, nursing involvement history, health status, and nursing workload from multiple perspectives. It automatically sends AI-generated reports, emergency contact information, daily nursing summaries, user status changes, suggestions for families, nursing workload reduction strategies, and family-oriented educational content to families. For example, in the event of an emergency such as a change in a user's physical condition or a fall, an immediate status report is sent to the family. It can also automatically generate weekly summaries of "this week's physical condition and emotional changes," "dietary intake trends," "sleep status," "recovery progress," and "words of gratitude to the family," and send them to the family. For example, if a family member requests to know about recent changes in their physical condition, a summary of their physical condition and emotional changes for the past week is automatically generated and sent to the family. For example, when a family member's nursing workload is high, the AI ​​automatically proposes suggestions such as "recommendation of nursing leave," "regional support service guidance," and "providing stress-related nursing information to families." Specifically, this Family Collaboration Department manages and stores family member profile information and communication settings in a dedicated database. The Family Collaboration Department receives structured data output from the Generative AI Processing Department. This system evaluates the importance score of the messages contained in this structured data. For urgent messages with an importance score exceeding a predetermined threshold, the system routes the message to an immediate delivery queue. The system uses SMS service or push notification protocols to send messages to family members' mobile devices with a latency of milliseconds. For regularly scheduled reports with an importance score below the threshold, the system stores the message in a batch processing queue. The system inputs accumulated data into the statistics module according to a specified schedule. This statistics module calculates the mean, variance, and slope of the trend line for vital sign data over the past week. The system inputs the calculated statistics and time-series data of sentiment analysis into a natural language generation model. This natural language generation model converts the statistical data into human-readable summary text. The system combines the summary text with image data generated from a charting library to dynamically generate an HTML report. The system sends the generated report to family members via email. The Family Collaboration Department collects access logs and message open rate data from family members' devices. This system inputs the data into a machine learning model to predict the optimal timing and granularity of notification for each family member. Based on the predictions, the system automatically updates notification settings. These processes by the Family Collaboration Department result in the selection of the optimal communication path commensurate with the importance of the information, and the technical effectiveness of data-driven personalized information delivery. This system is extremely useful in use cases such as alleviating family members' anxiety and facilitating smoother communication between the facility and families.Unlike existing systems that send uniformly formatted text messages, this system automatically generates rich content tailored to individual circumstances by combining dynamic data statistics and natural language generation technology. This system balances the optimization of communication resources with the improvement of user experience.

[0088] The Feedback Accumulation Department collects and manages various evaluation indicators, including feedback from families and staff, users' physical condition and emotional change history, nursing implementation results, emergency response history, workload data, AI proposal adoption rate, response speed, complaint rate, repeat response rate, family satisfaction, employee satisfaction, management evaluation, and external audit results. The Generative AI Processing Department utilizes this data for self-learning and model optimization, nursing procedure improvement, educational program optimization, business process improvement, risk management enhancement, and service quality improvement. For example, if families request that "recent reports are easy to understand" or "more detailed information is needed," the AI ​​automatically adjusts the report content and frequency. If staff provide feedback such as "AI proposals are not suitable for the situation" or "the proposals are accurate," the AI ​​automatically corrects the response generation algorithm to improve on-site adaptability. For example, if the adoption rate of AI proposals is low, the AI ​​automatically performs root cause analysis to improve the content, timing, and presentation of the proposals. For example, if the repeat response rate is high, the AI ​​analyzes past response history and automatically generates strategies to prevent recurrence. Specifically, this Feedback Accumulation Department functions as a data lake built on a distributed file system. In addition to explicit evaluation scores and freehand text sent from the terminal, this feedback accumulation unit also collects operation logs such as screen scrolling volume, dwell time, and click history as implicit feedback data. The system inputs the collected raw data into the data cleaning module. This module removes noise and imputes missing values. The system then inputs the cleaned data into the feature engineering pipeline. This pipeline calculates sentiment polarity scores from the feedback text and extracts the adoption / non-adoption of AI proposals as binary labels. The system constructs a training dataset from these feature pairs and labels. Using this training dataset, the system periodically performs fine-tuning of the neural network model for the generative AI processing unit. The system updates the model's weight parameters using backpropagation. The system calculates cross-entropy error or mean squared error as the loss function and minimizes the loss using gradient descent. The system incorporates a reinforcement learning framework, defining the adoption rate of AI proposals as the reward function. The system uses a proximal policy optimization algorithm to optimize the response generation strategy in a way that yields higher rewards. The system evaluates the performance of the updated model using a validation dataset. This system seamlessly deploys production models only after performance improvements have been confirmed. These processes, mediated by this feedback accumulation department, result in a continuous and self-disciplined improvement in the system's inference accuracy. This system is extremely useful in use cases such as self-evolving systems that adapt to field needs as their deployment period lengthens. Unlike existing systems that rely on manual rule updates and parameter tuning, this system automates data-driven machine learning pipelines. It prevents model staleness and consistently maintains optimal performance.

[0089] The Security Management Department employs multi-layered encryption to protect confidential data such as user information, nursing history, medical information, family information, and employee information, preventing unauthorized access and information leakage. For example, it uses quantum-resistant cryptography to protect communication paths, finely configures access permissions to restrict access to personal information, or records login and operation history as audit logs. The system is designed to send only the minimum necessary information to the terminal, displaying only response strategies and necessary historical information in a way that cannot be used to target specific individuals. For example, usernames are anonymized and displayed on employee terminals, while other user information is completely hidden on family member terminals, ensuring thorough privacy protection. Specifically, the Security Management Department implements an authentication and authorization control mechanism based on a zero-trust architecture. Even access from within the network boundary requires strict authentication for all requests. The Security Management Department utilizes public-key cryptography infrastructure to issue unique digital certificates to each terminal and user. The system uses the latest version of Transport Layer Security (TLS) protocols to establish end-to-end encrypted communication channels. Data stored in the database is transparently encrypted using Advanced Encryption Standard (AES) algorithms. The Security Management Department employs an attribute-based access control model. This system inputs attribute vectors such as the user's job title, department, current location, and access time into the evaluation engine. This evaluation engine uses logical operations to dynamically determine whether access permission is granted. The system applies a data anonymization module to the data payload sent to the terminal. This data anonymization module uses regular expression matching to detect patterns such as phone numbers and social security numbers within text data. The system replaces detected strings with irreversible characters such as asterisks. For image data, the system performs anonymization processing using generative adversarial networks to replace facial regions with synthetic, fictitious facial images. The security management department logs all access attempts and data manipulation events in a tampering detection log with a hash chain structure. These processes performed by the security management department result in a significant improvement in the system's resilience against internal crimes and external cyberattacks. This system is extremely useful in use cases requiring compliance with stringent requirements such as medical information protection laws. Unlike existing perimeter-based security models, this system implements a data-centric, multi-layered defense as an internal computer process. This system guarantees the highest level of confidentiality and integrity without compromising convenience.

[0090] The External Collaboration Department collaborates with multiple layers of external systems, including medical databases, emotion recognition engines, speech recognition engines, AI-powered psychological care support systems, vital sign sensors, family contact systems, disaster information systems, regional integrated support centers, administrative services, long-term care insurance systems, and external auditing systems, to acquire and utilize necessary information in real time. For example, in the case of voice input, the speech recognition engine transcribes the data into text, the emotion recognition engine determines the emotion, the department retrieves the latest treatment guidelines and medication information from medical databases, and obtains real-time temperature, blood pressure, and pulse data from vital sign sensors, reflecting this information in AI-based care proposals and risk assessments. For example, during disasters, the department retrieves information about affected areas from disaster information systems and automatically generates safety confirmations and evacuation guides for users and their families. For example, it collaborates with external auditing systems to automatically assess compliance with regulations and service quality, providing management with improvement suggestions. Specifically, this External Collaboration Department employs a service mesh architecture for communication between mediated microservices. For external application programming interfaces (APIs), this department uses RESTful or gRPC protocols to execute asynchronous remote procedure calls. To address response delays and temporary failures from external systems, this system implements a circuit breaker mode and an exponential backoff algorithm. It continuously sends PCM-formatted speech stream data in blocks to an external speech recognition engine. The system receives text tags and timestamp pairs returned from the speech recognition engine. This text data is then forwarded to an external emotion recognition engine, which outputs the emotional polarity of the text as a multi-dimensional vector. The external collaboration unit receives sensor data packets from vital sign sensor devices via the Bluetooth Low Energy protocol. The system decodes the payload within the data packets, extracting body temperature and blood pressure values ​​as floating-point data. These extracted values ​​are then written to a time-series database. The external collaboration unit receives JSON payloads of emergency alerts from a disaster information system via a webhook. The system calculates the semi-versus distance between the latitude and longitude coordinates within the payload and the coordinates of facilities registered in the system. If the calculated distance is within a predetermined radius, the system determines a high disaster risk. This determination is used as a trigger to queue evacuation guide generation tasks in the generative AI processing unit. These processes, performed through this external collaboration department, eliminate data silos between fragmented, heterogeneous systems, creating a unified data processing pipeline. This system is extremely useful in use cases such as rapid decision support in emergencies and automated acquisition of the latest medical insights. Unlike existing systems that rely on manual data transcription and batch processing, this system achieves real-time inter-system collaboration through an event-driven architecture. This system maximizes data freshness and accuracy.

[0091] The Terminal Interface Department operates on various terminals used by employees, users, and their families, providing multiple functions such as displaying AI proposals, editing response texts, inputting nursing records, acquiring vital sign data, sending image and voice data, confirming reports to family members, inputting feedback, viewing FAQs and manuals, listening to educational content, one-click notification in emergencies, managing business progress, and alarm notifications. For example, it can send AI-generated response texts on the chat screen with one click, or display response plans in real time during voice calls, or automatically pop up FAQs and manuals. For example, it can link with vital sign sensors to automatically perform AI analysis and issue alarms when abnormal values ​​are detected. For example, family members can use their smartphones to check the user's latest condition and receive AI-generated weekly summaries and emergency reports instantly. Specifically, this Terminal Interface Department provides a dynamic user interface built using a single-page application framework. This Terminal Interface Department utilizes virtual DOM technology to minimize screen redrawing and optimize rendering performance. This system receives inference result data in JSON format pushed from the server. This system stores the received data in local state management storage. The terminal interface unit detects updates to the state management storage and asynchronously updates only the relevant UI components. In voice call functionality, the system uses the WebRTC protocol to establish a peer-to-peer media stream. The system separates the voice track from the media stream and extracts the voice waveform data using the WebAudio API, which operates in a background thread. The system sends the extracted voice waveform data to the server at regular intervals. The system displays the real-time response text returned from the server overlaid in a floating window on the call screen. The terminal interface unit accesses the terminal's camera module to acquire image data. Before sending the image data, the system utilizes the terminal's GPU to perform image resizing and JPEG compression. This system achieves high-speed data transmission while minimizing network bandwidth consumption. The terminal interface unit temporarily buffers time-series data obtained from the vital signs sensor in the terminal's memory. The system applies a moving average filter to the buffered data to remove noise. If the filtered data exceeds a threshold, the system controls the terminal's vibration motor and LED indicator to generate a physical alarm. These processes performed by the terminal interface unit minimize the impact of network latency and provide users with a seamless and responsive operating experience. This system is extremely useful for rapid information input in busy nursing settings and for timely notification reception in emergencies. Unlike existing thin-client interfaces, this system implements a rich-client architecture that actively utilizes edge computing resources. This system distributes server load, improving overall system processing efficiency.

[0092] Thus, the system involved in this implementation includes a user information management department, an employee information management department, a generative AI processing department, a family collaboration department, a feedback accumulation department, a safety management department, an external collaboration department, and a terminal interface department. These departments work together to monitor each user's physical condition, emotional changes, and nursing history in real time from multiple perspectives, including family requests, employee status, environmental changes, and external risk information. It automatically generates and prompts optimal nursing actions and reports, provides employee support, family collaboration, risk management, educational support, and business improvement, achieving a highly sophisticated nursing support system that continuously performs self-learning, self-correction, environmental adaptation, and quality improvement. Generative AI plays an indispensable role in all processes, including consultation analysis, emotion recognition, response generation, self-learning, self-correction, environmental adaptation, family collaboration, employee education, trend analysis, automatic generation of nursing procedures, simultaneous notification in emergencies, automatic updating of medical guidelines, risk assessment, external collaboration, and legal compliance support. This makes it possible to achieve high-precision and flexible application in nursing settings and create multifaceted value, which is impossible with existing rule-based systems. Specifically, this system adopts an architecture that uses an event-driven message broker to loosely couple and link the various functional modules. This system uses a publish-subscribe model to asynchronously distribute state change events occurring in one module to all other interested modules. Based on this event stream, the system achieves high throughput, processing tens of thousands of transactions per second with zero latency. The system positions its generative AI processing unit as the central inference engine. The system aggregates multidimensional context vectors collected from various management units into the input layer of the generative AI processing unit. Through an autoregressive decoding process using a large-scale language model, the system probabilistically generates optimal action sequences for complex situations. The generated action sequences are input into an evaluation function, scored from a security or compliance perspective. If the score does not meet the benchmark, the system runs a self-correcting loop of dynamically adjusting prompts and performing re-inference. The system uses log data accumulated in the feedback accumulation unit to execute a continuous learning pipeline that combines offline batch learning and offline successive learning. The system adaptively updates the model's weight parameters in response to new and unknown situations and environmental changes. These processes result in the system functioning not merely as a data input / output device, but as a self-regulating, evolving intelligent agent. This system is extremely useful in the nursing industry, where staff shortages are severe, for use cases that simultaneously achieve automation of operations and standardization of care quality. Fundamentally different from existing systems that rely on predefined, static conditional branches, this system achieves highly sophisticated information processing based on probabilistic reasoning in high-dimensional spaces and dynamic parameter updates. This system expands the applicability of computer technology itself, contributing to the solution of complex real-world problems.

[0093] (Implementation steps) Step 1: Registration and updating of user, family member, and employee information The User Information Management Department stores and manages each user's nursing history, medical information, lifestyle habits and preferences, emotional and physical condition changes, and family contact history in a multi-layered and long-term manner. The Employee Information Management Department centrally manages each nursing employee's skill set, responsibility history, stress level, shift schedule, qualifications, past evaluations, workload, and educational history. The Family Collaboration Department manages each family member's contact information, desired notification methods, requests, feedback, and family meeting minutes. This information is continuously registered and updated via the Terminal Interface Department. For example, when a user reports a new allergy, an employee obtains a new qualification, or a family member changes their notification method, all information is reflected in real time. Specifically, this system receives HTTP POST or PUT requests from the Terminal Interface Department. The system parses the JSON-formatted registration data contained in the request payload. The system executes a schema validation module on the input data, checking for missing data types and required items, and string format errors. The system forwards the validated data to the database engines of each management department. In the relational database, the system initiates transaction processing. While ensuring ACID properties, the system atomically performs insert or update operations on multiple tables. When new allergy information is input, this system feeds the text data into a natural language processing (NLP) model. This NLP model, referencing a medical ontology dictionary, extracts standardized identification codes for the allergens. The system stores these extracted identification codes as foreign keys in the medical information table. When an employee's eligibility information is updated, the system recalculates the employee's profile vector. The updated profile vector is then overwritten and stored in the cache area of ​​the in-memory database. When the notification method for family members changes, the system dynamically rewrites the flags in the communication routing table. After the database update is complete, the system immediately publishes the change notification event to the message broker. These processes achieve the technical effect of maintaining data consistency and up-to-dateness across the distributed system in milliseconds. This system is extremely useful in high-concurrency use cases such as multiple employees updating information simultaneously. Unlike existing manual batch synchronization and file sharing, this system uses an event-driven architecture to implement real-time state synchronization as an internal computer process. This system mathematically eliminates data conflicts and inconsistencies.

[0094] Step 2: Input of daily nursing records, statements, vital signs, images, voice, and sensor information. Through the terminal interface, staff and users input and send various types of information, including daily nursing records (e.g., bathing care, medication management, food intake, excretion status, rehabilitation content), spoken content (e.g., "Headache today," "Thank you"), vital signs data (e.g., body temperature, blood pressure, pulse, blood oxygen), image data (e.g., photos of pressure sores, photos of food intake), voice data (e.g., speech, cough, breath sounds), and sensor information (e.g., fall sensors, bed exit sensors, door opening / closing sensors). For example, staff input nursing records using a tablet, users record spoken content using a voice input terminal, and vital signs sensors automatically send data. Specifically, the system initiates a dedicated data collection pipeline in the terminal interface to handle diverse input modalities. The system receives text-formatted nursing records from the tablet terminal and normalizes the character encoding to UTF-8. The system converts the analog voice signal acquired from the microphone array of the voice input terminal into digital voice waveform data with a sampling rate of 16kHz using an analog-to-digital converter. The system applies a noise reduction algorithm to this digital voice waveform data to attenuate background noise. This system continuously receives timestamped floating-point numerical sequences from vital signs sensors via a serial communication interface. It expands RGB image data acquired from the camera module into a tensor of pixel values ​​in memory. The system applies an affine transformation to this image tensor to correct image skew and size to a standard format. The system detects interrupt signals of binary status from bed exit and door opening / closing sensors at the edge gateway. This system encapsulates these heterogeneous data formats into a unified data packet with a common metadata header. The system uses a secure communication protocol to forward this unified data packet to the server-side receive buffer. These processes result in the efficient collection of both unstructured and structured data in a unified format. This system is extremely useful for automating multimodal information collection in the nursing field and for the integrated management of sensor networks. Unlike existing systems that rely on manual data transcription and individual system inputs, this system seamlessly integrates hardware-level signal processing and software-level data integration. The system minimizes the data input load and prevents information loss.

[0095] Step 3: Multi-angle analysis, emotion recognition, and abnormal sign detection based on generative AI. The generative AI processing unit performs multi-angle analysis on statements sent from the terminal, nursing records, vital sign data, image / voice / sensor information, and requests from family members. It accurately identifies changes in the user's emotions and physical condition, abnormal signs, disruptions in daily routines, environmental changes, and changes in family members' requests. For example, if the statement "My chest hurts a little today" is recorded, it is judged as "pain". By comprehensively referring to past medical history, medication information, vital sign data, past responses to chest pain, recent activity levels, diet, stress factors, and changes in temperature and air pressure, it automatically generates multiple nursing actions such as "contacting medical staff to recommend vital sign checks", "automatically sending a status report to the family", and "suggesting quiet instructions and enhanced observation to staff". Specific examples of prompts that generative AI can read include: "User ID U001 has a history of hypertension and is currently taking antihypertensive medication. Today's message is 'My chest hurts a little today.' Based on past nursing history and medical information, generate the necessary nursing actions and a report to the family." "User ID U002 recently said 'anxious.' Based on past emotional history and nursing records, generate a contact with psychological care staff and a reassuring conversation example." "Analyze pressure ulcer image data to generate progression assessment and response proposals." "Analyze the action history and environmental factors when the fall sensor is working to generate a fall cause analysis and recurrence prevention measures," etc. Specifically, this generative AI processing unit inputs the received integrated data packets into the multimodal parsing pipeline. The system inputs text data into a bidirectional encoder representation model, outputting a 768-dimensional text feature vector that considers the context. The system inputs image tensors into a pre-learned residual neural network, outputting a 2048-dimensional image feature vector. This system inputs a time-series array of vital sign data into a one-dimensional convolutional neural network, outputting a time-series feature vector that captures patterns of temporal variation. This feature vector is then input into a consolidation layer to generate a single unified context vector. This unified context vector is then input into an autoencoder model for anomaly detection. This autoencoder model compresses the input vector to a low-dimensional space and reconstructs it to its original dimension. The system calculates the reconstruction error between the input vector and the reconstructed vector. If the reconstruction error exceeds a predetermined threshold, the system determines that an unknown anomalous symptom exists. The system inputs text feature vectors into a fully connected layer for sentiment classification, applies the Softmax function, and outputs the probability distribution of sentiment categories such as "pain," "anxiety," and "joy." The system extracts the sentiment category with the highest probability as a label. The extracted sentiment labels, anomaly detection results, and historical data retrieved from the database are then input into a prompt generation module.This prompt generation module converts this data into structured JSON-formatted prompt strings. The system then inputs these prompt strings into a large-scale language model with billions of parameters. This large-scale language model uses a self-attention mechanism to compute dependencies between input token columns, autoregressively generating instructions for the optimal care action. These processes, based on this generative AI processing unit, result in the ability to integrate heterogeneous data in high-dimensional space, achieving a level of accuracy and speed in assessing situations that surpasses human cognitive capabilities. This system is extremely useful in use cases such as the early detection of complex symptoms and providing high-level decision support for inexperienced staff. Unlike previous systems that relied on simple threshold judgments or rule-based conditional branching, this system utilizes the nonlinear transformations of deep learning models to extract correlations between complex data. This system implements data-driven, high-level reasoning as internal computer processing.

[0096] Step 4: Automatic Generation and Terminal Prompts of Optimal Nursing Actions, Reports, and Instructions. Based on the analysis results, the generative AI processing unit automatically generates optimal nursing actions (e.g., contacting medical staff, checking vital signs, instructing for quiet, contacting psychological care staff, consulting a nutritionist, automatic reporting to family members, examples of conversations with staff, pressure ulcer care procedures, fall recurrence prevention strategies, infection control measures, rehabilitation plan reassessment, enhanced nighttime care, proposals for outreach support, recommendations for recreational activities, and generation of educational content for family members), and provides real-time prompts to staff and family terminals. For example, the staff terminal might display "Today, someone complained of chest pain. Please contact medical staff and perform a vital sign check," while the family terminal automatically receives a report such as "Today, due to a complaint of chest pain, medical staff are responding." Specifically, the system inputs the text stream output from the large-scale language model into the natural language understanding module. This module extracts the intent, target, and execution time of the action from the text. The system converts the extracted information into a structured task object. This task object is then input into the task scheduling engine. This task scheduling engine calculates the urgency and importance scores for each task. Tasks with urgency scores exceeding a threshold are inserted into a priority queue. Tasks are retrieved from the priority queue, and the terminal identifier of the employee is identified. Task data is sent to the identified terminal identifier using a push notification protocol. The received task data is parsed in the employee's front-end application. Based on the task's urgency, the background color of UI components on the screen is changed to red, and a warning audio file is played. When a report task object for family members is generated, it is input into the message format conversion module. This module performs dictionary lookup to replace technical terms with common expressions. The converted text data is sent to the family member's terminal via an SMS gateway or email server. The system receives a completion status code and updates the sending history table in the database. These processes enable the distribution of generated inference results as immediately actionable actions to relevant personnel. This system is extremely useful in use cases such as rapid information sharing in emergencies and preventing task omissions. Unlike previous systems that relied on human-based communication networks or verbal instructions, this system autonomously routes and prioritizes information. It reduces communication latency to the millisecond level, dramatically improving the determinism of business processes.

[0097] Step 5: Staff and family members respond and provide feedback. Staff and family members refer to the AI-provided nursing actions and report content to respond in practice, inputting response results, additional information, and feedback (e.g., "The AI ​​proposal is accurate," "Does not match the situation," "Family report is easy to understand," "Want more detailed information," etc.) from the terminal interface. For example, staff record "Vital signs check performed," "Reported to doctor," while family members input "Satisfied with report content," "Want to know more about sleep status." Specifically, the system renders a dynamic input form on the terminal interface. The system provides a checkbox component to manage the execution status of the prompted nursing actions. The system detects touch events when the user clicks a checkbox. Using this event as a trigger, the system generates a status update payload containing the task completion time and user identifier. The system uses an asynchronous communication protocol to send this payload to the server. For feedback input, the system provides a 5-level star rating component and a free-writing text area component. The system inputs the user's text data into the client-side validation function in real time. The system detects and blocks attempts to inject illegal characters or scripts. This system serializes cleaned text data and evaluation scores into a JSON object. This JSON object is then sent to the server-side feedback accumulation unit. The system also allows for voice-based feedback. Voice streams from the terminal's microphone are input into the edge-side speech recognition model and transcribed into text. The transcribed feedback data is then sent to the server. These processing methods result in the efficient collection of explicit user evaluation data in a structured format. This system is extremely useful for use cases such as evaluating system usefulness and collecting learning data for the next model iteration. Unlike previous questionnaires conducted later, this system utilizes an internal computer processing mechanism to obtain feedback immediately after task execution while maintaining context. This system ensures data freshness and accuracy, enabling seamless data delivery to the machine learning pipeline.

[0098] Step 6: Feedback, Historical Data Accumulation, and AI Self-Learning and Optimization. The Feedback Accumulation Department accumulates various evaluation indicators, including feedback from families and staff, users' physical condition and emotional changes history, nursing implementation results, emergency response history, workload data, AI proposal adoption rate, response speed, complaint rate, re-response rate, family satisfaction, employee satisfaction, management evaluation, and external audit results. The Generative AI Processing Department utilizes this data for self-learning, model optimization, nursing procedure improvement, educational program optimization, business process improvement, risk management enhancement, and service quality improvement. For example, when the adoption rate of AI proposals is low, the AI ​​automatically performs cause analysis and improves the content, timing, and expression of the proposals. Specifically, this system loads the collected feedback data and various log data into the fact table of the distributed data warehouse. This system periodically executes the extraction, transformation, and loading processing pipeline. This system extracts feature vectors from the raw data to construct a dataset for machine learning. This system constructs a supervised learning model with the AI ​​proposal adoption rate as the target variable and the input context vector as the explanatory variable. This system uses a gradient boosting decision tree algorithm to calculate the importance of features that contribute to a decrease in adoption rate. By identifying highly important features, the system statistically analyzes the root causes of proposal rejection. Based on this analysis, the system automatically begins fine-tuning the large-scale language model for the generative AI processing unit. The system uses pairs of prompts and outputs for adopted proposals as positive examples and pairs of unadopted proposals as negative examples. The system applies reinforcement learning techniques based on human feedback. It trains a reward model and uses this model to update the parameters of the large-scale language model using a proximal policy optimization algorithm. The system uses a reserved test dataset to evaluate the inference accuracy of the updated model. The system automatically replaces the production model container in a rolling update manner only if the BLEU score or ROUGE score, used as an evaluation metric, exceeds a pre-defined baseline. These processes result in a continuous improvement in the system's inference performance without human intervention. This system is extremely useful in use cases such as building self-evolving systems that dynamically adapt to changes in the environment or user preferences. Unlike previous static software, this system uses a data-driven, self-optimizing loop as internal computer processing. This prevents system obsolescence and maximizes its long-term value.

[0099] Step 7: Implementation of External Collaboration, Security Management, and Legal Compliance. The External Collaboration Department collaborates with medical databases, emotion recognition engines, speech recognition engines, psychological nursing support AI, vital sign sensors, family contact systems, disaster information systems, regional integrated support centers, administrative services, nursing insurance systems, and external auditing systems at multiple levels. This allows for real-time acquisition and utilization of necessary information. The Security Management Department implements multi-layered encryption for confidential data such as user information, nursing history, medical information, family information, and employee information to prevent unauthorized access and information leakage, thoroughly implementing access control management, audit log recording, privacy protection, and legal compliance. Specifically, this system operates the API gateway within the External Collaboration Department. This API gateway proxies all outbound and inbound communications to external systems. The system parses the header information of communication data packets and verifies the validity of access tokens based on the OAuth 2.0 protocol. The system only routes requests to backend microservices if the token is valid. The system issues GraphQL queries to external medical databases, selectively retrieving only the required data fields. The system temporarily stores the retrieved data in a memory cache to accelerate subsequent inference processing. This system applies transparent data encryption to all database transactions within the security management department. It uses the AES-256 algorithm to encrypt data blocks on memory at the physical level. A dedicated hardware security module manages the lifecycle of encryption keys. A role-based access control module is applied to user access requests. The system compares the user's permission level with the security label of the requested resource, using bitwise operations to determine whether access is permitted. All system calls, API requests, and database queries are recorded as audit logs along with timestamps and user identifiers. These audit logs are archived in a secure object store with Write Once Read Many properties. The system periodically anchors the hash values ​​of the audit logs to a blockchain network, mathematically making post-transaction tampering impossible. These processes result in a technical effect that balances secure data exchange with external systems with rigorous protection of internal data. This system is extremely useful in use cases requiring stringent compliance with laws such as HIPAA in healthcare information systems. Unlike previous add-on security measures, this system integrates security and auditing mechanisms into its fundamental architecture design. This system minimizes the risk of information leakage and guarantees the system's reliability.

[0100] (specific use case) For example, in a nursing facility, if a user says "My chest hurts a little today" after a morning vital signs check, the terminal interface department simultaneously sends this statement along with vital signs data such as temperature, blood pressure, and pulse to the server. The user information management department immediately extracts relevant data such as the user's medical history (hypertension and heart disease), medication dosage (antihypertensive medication), past reports of chest pain, recent sleep patterns and dietary intake, stress factors, and changes in air pressure. The employee information management department extracts information such as the medical and nursing skills of the employee responsible for the user, current workload, stress level, and recent shift schedule. The generative AI processing department comprehensively analyzes this information, identifies the emotion of "pain" from the statement, and automatically generates multiple nursing actions such as "contacting medical staff to re-check vital signs," "automatically sending a status report to the family," and "suggesting quiet instructions and enhanced observation to the responsible staff." The system displays a message on the employee terminal: "The user complains of chest pain. Please contact medical staff and have vital signs checked again. Please remain quiet and report any abnormalities immediately." It also automatically sends a report to the family terminal: "Today, due to complaining of chest pain, medical staff are responding. Updates will be provided as needed." Specifically, the system uses a speech recognition module to convert the user's analog speech signal into digital text data. This system synchronizes the converted text data with vital sign data acquired from various sensors. The system structures the vital sign data into multi-dimensional time-series numerical vectors. The system issues a database query to the user information management department. The system extracts the user's past medical history and medication history as natural language text. The system obtains the skill level and current task count of the responsible staff from the employee information management department as numerical data. The system integrates this heterogeneous data and inputs it into the generative AI processing unit. This generative AI processing unit possesses a large-scale language model based on the Transformer architecture. The system tokenizes the input data and converts it into high-dimensional embedding vectors. For example, the system generates a 512-dimensional input tensor. This system, as a preprocessing step for input data, performs normalization of vital sign data and interpolation of missing values. As a specific example of input data, this system provides the model with the text "My chest hurts a little today," numerical vectors of systolic blood pressure of 150 and heart rate of 90, and categorical data such as past medical history markers. The model uses a self-attention mechanism to calculate the hidden correlations between the input elements. As output, this model generates a probability distribution for emotion classification and structured data for recommended nursing actions. As a specific example of output data, this model outputs a pain probability of 0.95 as an emotion score, a level 4 as an urgency score, and a generated text string. The system compares the output urgency score with a pre-set threshold.This system immediately initiates an interruption process when the score exceeds a threshold. The system forwards the generated text string to the employee's terminal display module. Simultaneously, the system invokes an external communication application programming interface (API). The system sends push notifications to family members' terminals. Through these processes, the system achieves objective and rapid triage without waiting for human judgment. As a result, the system effectively prevents delays in medical intervention and dramatically improves user safety. This system is extremely useful in healthcare facilities and hospitals.

[0101] In addition, when a user states "I haven't been able to sleep at night lately," the User Information Management Department extracts past sleep records, medication history, daytime activity levels, caffeine intake, nighttime urination frequency, noise levels, lighting environment, and past responses to insomnia complaints. The Generative AI Processing Department automatically generates suggestions such as "Consult a doctor," "Report sleep status to family members," and "Instruct staff to strengthen nighttime care and adjust the environment." The employee terminal displays "User complains of insomnia. Please strengthen nighttime care and adjust lighting and noise levels," and the family terminal automatically reports "Recently, nighttime sleep seems to be lighter. We plan to consult a doctor." Specifically, the system inputs the user's text into a natural language processing engine. The system categorizes the intent of the statement as a complaint of insomnia. The system extracts time-series data from the User Information Management Department, including sleep duration and awakening frequency over the past month. The system acquires noise level and illuminance data from environmental sensors as time-series arrays. The system inputs this time-series data into a recurrent neural network or long short-term memory network. The system extracts feature vectors representing time dependencies from the time-series data. This system combines the extracted feature vectors with the embedded vectors of the spoken text. The combined vector is then input into the multilayer perceptron of the generative AI processing unit. As preprocessing of the input data, the system performs moving averages and outlier removal on the time-series data. As a specific example of input data, the system provides the model with values ​​such as the average sleep time of four hours over the past seven days, three nighttime awakenings, and an average indoor illuminance of fifty lux. The model repeatedly applies nonlinear transformations to the input vectors. As output, the model generates classification labels for the main factors presumed to cause insomnia and text tags representing coping strategies. As a specific example of output data, the model outputs an environmental factor probability of 0.8, environmental adjustment instructions as recommended action codes, and a generated report. The system parses the output recommended action codes. Based on the parsing results, the system sends control signals to the facility's environmental control modules to indicate dimming of lighting and adjustment of air conditioning. The system renders the generated report on the user interfaces of employee and family terminals. Through these processes, the system goes beyond simple event recording and automatically executes data-driven preventative care. As a result, this system significantly reduces the cognitive load required for staff to monitor the situation and equalizes the quality of care. This system can be applied to areas such as elderly care facilities and home care support services.

[0102] Furthermore, in the context of pressure ulcer care, when employees send photos of the pressure ulcer site from their terminals, the generative AI processing unit uses image analysis AI to automatically determine the progression of the pressure ulcer and automatically generate messages such as "Recommend reporting to the doctor," "Dressing change instruction," and "Report the progress to the family." The employee terminal displays "The pressure ulcer is progressing upwards. Please report to the doctor and change the dressing," while the family terminal automatically reports "The doctor is confirming the progress of the pressure ulcer site." Specifically, this system receives two-dimensional color image data acquired from the terminal's camera module. This system converts the image data into a three-dimensional tensor format. This system uses a convolutional neural network as the image analysis AI. This system performs resizing and pixel value normalization on the input image. This system applies image rotation and inversion as data augmentation techniques to improve the model's generalization performance. As a specific example of input data, this system feeds a 256x256 pixel, three-channel red-green-blue image tensor to the model. This model extracts local and global spatial features from the image through multiple convolutional and pooling layers. This model uses extracted features to classify the severity of pressure ulcers. As output, the model generates probability distributions corresponding to each stage of progression, as well as segmentation masks representing lesion sites. Specific examples of the output data include a probability of 0.85 for stage three, a 20% area ratio of necrotic tissue, and lesion boundary coordinates. The system selects the maximum value of the output probability distribution to determine the current progression. The system compares the determined progression with past assessments. When the system determines that the progression is worsening, it dynamically constructs and inputs prompts to encourage emergency response into the text generation model. The system pushes and distributes the instructions output from the text generation model to employee terminals. Through these processes, the system eliminates subjective evaluations based on human visual perception. As a result, the system achieves objective and high-precision condition assessment based on pixel-level computation in high-dimensional space, ensuring that appropriate intervention opportunities are not missed. This system is highly effective in use cases such as telemedicine and access care.

[0103] In addition, when the fall sensors are activated, the generative AI processing unit analyzes recent action history and environmental factors, and immediately executes actions such as "fall cause analysis," "recurrence prevention strategy proposals," and "automatic reporting to family members, doctors, and managers." The employee terminal displays "A fall has occurred. Please conduct on-site verification and observe the user's condition, and discuss recurrence prevention strategies," and the family terminal automatically reports "A fall occurred today, but it is not serious. We are discussing future recurrence prevention strategies." Specifically, this system receives high-frequency time-series signals from three-axis accelerometers and gyroscopes installed in the facility. The system detects abrupt changes in sensor signals and generates a hardware interrupt. The system extracts sensor data from a certain period immediately preceding the interrupt as a multi-dimensional time-series tensor. Simultaneously, the system acquires 3D point cloud data from a depth camera and the status of the bed surface moisture detection sensors as environmental factor vectors. The system integrates this multimodal data and inputs it into the anomaly detection model of the generative AI processing unit. As preprocessing of the input data, the system performs sensor noise filtering and point cloud data downsampling. This system, as a specific example of input data, feeds the model with acceleration vectors sampled at 100 Hz over ten seconds, the true / false values ​​of bed wetness indicators, and the user's recent walking speed data. The model uses a spatiotemporal attention mechanism to identify abnormal patterns in movement leading up to the fall. As output, the model generates a classification score representing the root cause of the fall and a structured text describing recurrence prevention measures. As a specific example of output data, the model outputs a slip-and-fall probability of 0.9, recommended bed cleaning instructions as environmental improvement measures, and an automatically generated draft incident report. Based on the output classification score, the system executes an application programming interface (API) call to the emergency notification system. The system stores the generated report in a database and distributes it instantly to the terminals of relevant personnel. Through these processes, the system not only accelerates post-incident response but also automates the initiation of data-driven, logical recurrence prevention measures. As a result, the system achieves the technical effect of improving the overall safety management level of facilities and mathematically reducing the risk of future accidents. This system is applied in environments such as elderly care facilities and rehabilitation hospitals.

[0104] As specific examples of prompts that generative AI can read, consider: "User ID U001 has a history of hypertension and is currently taking antihypertensive medication. Today's statement is 'My chest hurts a little today.' Based on past nursing history and medical information, generate the necessary nursing actions and a report to family members." "User ID U002 recently stated 'I can't sleep at night.' Based on past sleep records, medication history, daytime activity levels, and environmental factors, generate a consultation with a doctor, a report to family members, and instructions for environmental adjustments to staff." "Analyze pressure ulcer image data to generate progression assessment and response proposals." "Analyze the action history and environmental factors when the fall sensor is activated to generate a fall cause analysis and recurrence prevention strategies." Specifically, this system has a dynamic prompt generation module. This system performs the process of embedding structured data extracted from the database into templates of natural language. This system does not treat prompts merely as strings, but uses specific tags or markup languages ​​for formatting so that the model can accurately interpret the context. This system inputs the formatted prompts into a word segmenter. This system converts the strings into a sequence of integer indices and further maps them to an embedded representation on a continuous vector space. This system, as a preprocessing step for input data, removes unnecessary stop words and adjusts the weights of important words and phrases. As a concrete example of input data, it provides the model with a multi-dimensional tag vector containing system prompts, user prompts, and contextual data. The system generates an attention mask to improve computational efficiency by skipping the computation of filler tags. Based on the input prompt vector, the model successively calculates the probability distribution of the next word. As output, the model outputs a sequence of generated text tags and a log-likelihood representing the generation probability of each tag. As a concrete example of output data, the model outputs a list of nursing actions, a report written in a polite, family-oriented style, and an output confidence score. The system uses a decoder to inversely convert the output tag sequence into a human-readable string. The system performs regular expression validation on the output text. In cases where essential items are missing, the system runs a feedback loop requiring the model to regenerate. Through these processes, the system suppresses the illusions inherent in large-scale language models. As a result, the system demonstrates a technological ability to guarantee the stability and accuracy of outputs, achieving reliable information generation in the critical field of healthcare and nursing. This system is widely used in fields such as automated response systems or business support assistants.

[0105] In this way, the system of the present invention collects and analyzes various data such as user speech, vital signs, images, and sensor information in real time. Generative AI automatically generates optimal nursing actions, reports, and instructions, and immediately prompts and sends them to staff and family terminals, thereby significantly improving the safety, efficiency, family cooperation, and service quality of the nursing environment. Specifically, this system adopts a distributed computing architecture for comprehensively processing multimodal data. The system inputs heterogeneous data such as text, numerical values, images, and time-series signals into a dedicated encoder network. The system projects the feature vectors output from each encoder into a common latent space. The system combines the projected vectors to perform comprehensive inference. The system performs computationally intensive deep learning model inference processing on a parallel computing cluster using image processing devices. The system performs lightweight preprocessing and threshold determination on the edge device side. The system saves overall system communication bandwidth and minimizes end-to-end latency through cloud and edge collaborative processing. The system has a continuous learning pipeline. The system accumulates the history of staff correcting or acknowledging AI proposals on the terminal as real data. This system uses accumulated data to periodically fine-tune the model's weight parameters. It employs backpropagation to update parameters in the direction that minimizes the loss function. Through these processes, the longer the system is used, the more accurate its reasoning becomes, adapting to the facility's inherent operational rules and the individual characteristics of its users. As a result, this system completely replaces existing business processes that rely on human experience and intuition with data-driven and self-evolving computer processing. It balances efficient use of computing resources with increased processing speed, leveraging the technical benefits of improved computer technology itself. This system functions as a core information foundation in smart hospitals or next-generation nursing facilities.

[0106] 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.

[0107] Data generation model 58 is what is known as generative AI (Artificial Intelligence).

[0108] One example of data generation model 58 includes ChatGPT (registered trademark) (Internet search).<URL: https: / / openai.com / blog / chatgpt> Generative AI, etc.

[0109] 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.

[0110] 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 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. Additionally, 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.

[0111] 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.

[0112] [Second Implementation] Figure 3An example of the configuration of the data processing system 210 according to the second embodiment is shown.

[0113] 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.

[0114] 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).

[0115] 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.

[0116] 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.

[0117] 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).

[0118] 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.

[0119] 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.

[0120] 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.

[0121] 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).

[0122] 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.

[0123] 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".

[0124] (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.

[0125] (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.

[0126] 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.

[0127] 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.

[0128] 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.

[0129] [Third Implementation] Figure 5 An example of the configuration of the data processing system 310 according to the third embodiment is shown.

[0130] 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.

[0131] 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).

[0132] 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.

[0133] 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.

[0134] 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).

[0135] 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.

[0136] 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.

[0137] 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.

[0138] 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 special processing unit 290.

[0139] In the head-mounted terminal 314, the processor 46 performs the receiving and output processing.

[0140] 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.

[0141] 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".

[0142] (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.

[0143] (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.

[0144] 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.

[0145] 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.

[0146] 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.

[0147] [Fourth Implementation] Figure 7 An example of the configuration of the data processing system 410 according to the fourth embodiment is shown.

[0148] 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.

[0149] 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).

[0150] 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.

[0151] 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.

[0152] 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).

[0153] 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.

[0154] 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.

[0155] 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.

[0156] 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.

[0157] 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 special processing unit 290.

[0158] 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.

[0159] 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".

[0160] (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.

[0161] (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.

[0162] 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.

[0163] 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.

[0164] 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.

[0165] 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 system determines the user's emotions. 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.

[0166] 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.

[0167] 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.

[0168] 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).

[0169] 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.

[0170] 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."

[0171] 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.

[0172] The above description primarily focuses on the functions of the data processing device 12 to illustrate the system disclosed herein. However, the system disclosed herein may not necessarily be implemented on a server. The system disclosed herein can be implemented as a general information processing system. For example, it can be implemented as a software program running on a personal computer, an application running on a smartphone, etc. The methods disclosed herein can be provided to users in the form of SaaS (Software as a Service).

[0173] In the above embodiments, an example of specific processing performed by a single computer 22 is given. However, the technology disclosed herein is not limited to this, and distributed processing of specific processes by multiple computers, including computer 22, is also possible. For example, the data generation model 58 may also be provided in an external device of the data processing apparatus 12, where data is generated based on input data.

[0174] In the above embodiments, an example of storing a specific processing program 56 in memory 32 is illustrated, but the technology disclosed herein is not limited thereto. For example, the specific processing program 56 may also be stored in a portable computer-readable non-transitory storage medium such as a USB (Universal Serial Bus) memory. The specific processing program 56 stored in the non-transitory storage medium is installed in the computer 22 of the data processing device 12. The processor 28 performs specific processing according to the specific processing program 56.

[0175] Alternatively, the specific processing program 56 can be stored in a storage device such as a server connected to the data processing device 12 via a network 54, and the specific processing program 56 can be downloaded and installed into the computer 22 upon request from the data processing device 12.

[0176] Furthermore, it is not necessary to store the entire specific process 56 in a storage device such as a server connected to the data processing device 12 via the network 54, nor is it necessary to store the entire specific process 56 in the memory 32; a portion of the specific process 56 may also be stored.

[0177] As a hardware resource for performing specific processing, various processors can be used. For example, a CPU can be listed as a processor; it is a general-purpose processor that functions as a hardware resource for performing specific processing by executing software, i.e., a program. Furthermore, as a processor, a dedicated circuit can be listed as a processor; it is a processor with a circuit structure specifically designed for performing specific processing, such as an FPGA (Field-Programmable Gate Array), a PLD (Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit).

[0178] Every processor has built-in or connected memory, and every processor performs specific processes by using memory.

[0179] The hardware resources for performing a specific process can consist of one of these various processors, or a combination of two or more processors of the same or different types (e.g., a combination of multiple FPGAs, or a combination of a CPU and an FPGA). Furthermore, the hardware resources for performing a specific process can also be a single processor.

[0180] As an example of a single processor, the first type consists of a combination of one or more CPUs and software, which functions as a hardware resource to perform specific processing. The second type uses a processor, such as a System-on-a-chip (SoC), which implements the entire system functionality, including multiple hardware resources performing specific processing, using a single IC chip. In this case, the specific processing is implemented using one or more of the aforementioned processors that serve as hardware resources.

[0181] Furthermore, as the hardware architecture of these various processors, more specifically, circuits combining semiconductor elements and other circuit components can be used. Moreover, the specific process described above is merely an example. Therefore, it goes without saying that, without departing from the main point, unnecessary steps can be removed, new steps can be added, or the processing order can be changed.

[0182] The foregoing descriptions and illustrations are detailed explanations of the parts covered by this disclosure and are merely one example of this disclosure. For instance, the descriptions of the above-described structure, function, role, and effect are just one example of the structure, function, role, and effect of the parts covered by this disclosure. Therefore, it goes without saying that, without departing from the spirit of this disclosure, unnecessary parts can be deleted, new elements can be added, or replacements can be made to the foregoing descriptions and illustrations. Furthermore, to avoid confusion and facilitate understanding of the parts covered by this disclosure, explanations of technical common sense that does not require special explanation for implementing this disclosure have been omitted from the foregoing descriptions and illustrations.

[0183] All documents, patent applications and technical standards described in this specification are incorporated herein by reference as if they were specifically and individually described as incorporated by reference.

[0184] In addition, the following notes further explain the above explanation.

[0185] (Note 1) A nursing support system using generative AI includes a user information management department, an employee information management department, and a generative AI processing department, characterized in that... The user information management department stores and manages each user's nursing history, medical information, living habits, preferences, history of changes in emotions and physical condition, family contact history, etc., and collects and extracts various information from users' statements and log records, vital sign data, requests from family members, etc. in real time. The Employee Information Management Department centrally manages each nursing employee's skill set, responsibility history, stress level, shift schedule, qualification information, past evaluations, workload, and educational history, and dynamically allocates workload and adjusts workload based on the employee's operational status, adaptability, and stress level; and The generative AI processing unit uses generative AI to analyze the user's speech and recorded content in natural language, accurately identifying changes in emotions and physical condition. While comprehensively referencing various information such as the user's medical information and nursing history, the staff's skills and stress levels, and the family's requests, it automatically generates optimal nursing actions, emergency responses, automatic reports to family members, specific instructions to staff, and psychological care proposals in real time. Furthermore, it continuously performs self-learning based on feedback from family members and staff obtained after the visit, autonomously optimizing model accuracy and response content. At the same time, it achieves personalized optimized care for each user, rapid response in emergencies, reduced staff workload, and enhanced family collaboration.

[0186] (Note 2) The nursing support system using generative AI as described in Appendix 1 is characterized in that, When analyzing user statements and log records, vital sign data, and requests from family members, the generative AI processing unit works in conjunction with external emotion recognition engines, medical databases, psychological nursing support AI, and speech recognition engines to accurately determine changes in the user's emotions and physical condition. Based on the determination results, it generates multiple response strategies in real time, such as contacting medical staff, recommending vital sign checks, contacting psychological nursing staff, automatically reporting to family members, providing specific conversation examples to staff, and proposing nursing steps. Furthermore, it generates personalized and optimized responses that conform to the context, referring to the user's past nursing history and medical information, as well as the family's requests, and provides specific response instructions and reports to staff terminals and family terminals.

[0187] (Note 3) The nursing support system using generative AI as described in Appendix 1 is characterized in that, The generative AI processing unit accumulates feedback from family members and staff after receiving patients, the history of users' physical condition and emotional changes, nursing implementation results, emergency response history, workload data, etc. Using this feedback data, new nursing requirements, and environmental changes, it continuously optimizes the parameters and response generation algorithms of the generative AI model in a self-learning manner. When new nursing patterns or changes in medical conditions are detected, it autonomously generates coping strategies for the new patterns using clustering or pattern recognition technologies. When problems occur, it automatically performs readjustment or relearning of model parameters through a self-correction function.

Claims

1. A contact center system using generative AI, comprising a customer information management department, an operator information management department, and a generative AI processing department, characterized in that, The customer information management department stores each customer's consultation history, attribute information, past reception results, customer preferences, and channel information used in the long term, and can refer to the customer's past history and attribute information in real time when a consultation occurs. The operator information management department manages information such as each operator's skill set, responsibility history, stress level, and scope of business they can handle, and dynamically allocates and controls resources based on the operator's operational status and adaptability. as well as The generative AI processing unit uses generative AI to analyze customer inquiries in natural language, identifies emotions in the inquiry text in real time, and automatically generates the best response strategy while comprehensively considering information such as the customer's past history and attributes, the operator's skills and stress level. It also continuously performs self-learning using customer satisfaction and feedback information obtained after customer reception, autonomously optimizing model accuracy and response content. It can adapt to changes in consultation models and market environment, and autonomously readjust model parameters through self-correction function when problems occur, thereby improving customer satisfaction, increasing business efficiency, and reducing operator turnover.

2. The contact center system using generative AI as described in claim 1, characterized in that, When performing emotion recognition processing on consultation content, the generative AI processing unit uses an external emotion recognition engine or deep learning model to accurately determine the customer's emotions. Based on the determination results, it generates multiple response strategies such as apologies, explanations, proposals, and additional support. Furthermore, it generates personalized and optimized responses that conform to the context in real time by referring to the customer's past consultation history and attribute information, and prompts the operator with specific response instructions or proposals.

3. The contact center system using generative AI as described in claim 1, characterized in that, The generative AI processing unit accumulates customer satisfaction scores and operator feedback information obtained after customer reception. Using this feedback data, it continuously optimizes the parameters and response generation algorithm of the generative AI model in a self-learning manner. When new consultation patterns or changes in the market environment are detected, it autonomously generates response strategies for new patterns using clustering or pattern recognition technologies. When problems occur, it automatically performs readjustment or relearning of model parameters through a self-correction function.

Images