An intelligent AI credit assistant system based on a large language model

Abstract

The application relates to an intelligent AI credit assistant system based on a large language model, which comprises a mobile phone terminal AI assistant module, an intention recognition and task arrangement module, an intelligent information collection module, an end-cloud collaborative risk control module and an intelligent post-loan management module; the mobile phone terminal AI assistant module parses user intention through a natural language understanding engine; the intention recognition and task arrangement module automatically generates a task execution chain; the intelligent information collection module adopts a double-channel strategy of ''active collection + dialogue completion'' to extract structured data from a mobile phone terminal legal information source; the end-cloud collaborative risk control module comprises end-side pre-evaluation and cloud-side fine evaluation; the end side uses a light-weight risk control model for preliminary screening, and the cloud side uses a comprehensive scoring card model, a deep learning risk control model and a rule engine for multidimensional evaluation; and the intelligent post-loan management module generates personalized repayment reminders based on user portraits; the application improves the process efficiency and risk control precision of credit business, and realizes front-end risk control and real-time feedback.

Description

Technical Field

[0001] This invention relates to the fields of financial technology and artificial intelligence, and in particular to an intelligent AI credit assistant system based on a large language model. Background Technology

[0002] In the traditional lending process, users need to open a mobile app, manually fill in their personal information (such as identity information, proof of income, employment information, etc.), then choose a suitable product from multiple lending options, submit the application, wait for the risk control system to evaluate it, and finally obtain the approved credit limit. In the post-loan management stage, users also need to open the app regularly to make repayments. The entire process involves multiple steps and page jumps, making it lengthy and user-unfriendly, resulting in a poor user experience and high churn rate.

[0003] Currently, industry improvement solutions mainly focus on optimizing app interface interaction and simplifying form filling. However, these solutions still rely on users' active operation and cannot fundamentally solve the problem of lengthy processes. Some financial institutions have introduced simple chatbots for customer service Q&A, but these chatbots are based on rules or simple answer template matching and cannot understand complex user intentions, let alone replace users in completing core business operations such as loan applications, risk assessments, and repayment management. In addition, traditional risk control systems are independent of user interfaces, and risk control results cannot be fed back to users in real time. Users lack transparency and guidance while waiting for approval.

[0004] While large language models have made significant progress in the field of natural language processing, existing technologies for their deep integration with the entire credit business process are still immature, mainly due to the following technical problems: The generalized large language model lacks knowledge of the credit domain and has difficulty accurately understanding the professional terminology and user intent related to credit business. The existing risk control system adopts a centralized cloud assessment model, where users submit applications and wait for batch processing in the background, which cannot provide real-time feedback. Information collection during the credit application process relies on manual input by users, which is inefficient and prone to errors; Post-loan management adopts a passive strategy and cannot proactively provide personalized services based on individual user circumstances. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent AI credit assistant system based on a large language model, which solves the technical problems in the prior art such as lengthy credit business processes, delayed risk control feedback, low information collection efficiency, and passive post-loan management.

[0006] To achieve the above objectives, the present invention provides an intelligent AI credit assistant system based on a large language model, comprising a mobile AI assistant module, an intent recognition and task orchestration module, an intelligent information collection module, an end-to-cloud collaborative risk control module, and an intelligent post-loan management module. The mobile AI assistant module is deployed on the user's mobile terminal and has a built-in lightweight large language model that has been fine-tuned in the field of credit business. Users interact with the mobile AI assistant module through natural language. The intent recognition and task orchestration module is deployed on a cloud server and transforms the user's natural language requirements into a sequence of executable business tasks. The intelligent information collection module adopts a dual-channel strategy of active collection and dialogue completion to obtain user information; The edge-cloud collaborative risk control module divides the risk control assessment into two stages: edge-side pre-assessment and cloud-based fine assessment, realizing risk control in advance and real-time feedback. The edge-cloud collaborative risk control module includes an edge-side pre-assessment unit and a cloud-based fine assessment unit. The edge-side pre-assessment unit includes a lightweight risk control model deployed on a mobile device. It adopts a lightweight model based on gradient boosting decision trees or a neural network model compressed by knowledge distillation, and performs preliminary risk screening based on locally collected user behavior data. The cloud-based assessment unit, deployed on a remote server, includes a credit scoring model, an AI risk control model, a rules engine, and a fusion decision-maker. The credit scoring model uses logistic regression and linear regression algorithms to score credit based on user credit history, income level, debt ratio, and job stability. The AI ​​risk control model uses deep neural networks and graph neural networks to identify potential risk patterns from multi-dimensional features. The rules engine performs mandatory regulatory compliance checks. The fusion decision-maker uses weighted voting and stacking ensemble learning methods to weightedly fuse the assessment results from the credit scoring model, AI risk control model, and rules engine to generate the final risk control decision. The intelligent post-loan management module is used to perform proactive post-loan management.

[0007] Preferably, the mobile AI assistant module serves as the sole entry point for user interaction with the credit system, supporting both voice and text input methods. The mobile AI assistant module includes: The natural language understanding engine is used to parse the semantics of user input. It adopts an encoder-decoder model based on the Transformer architecture to perform word segmentation, named entity recognition, and semantic role labeling on the user's natural language input, and extract credit-related entity information. The dialogue state manager adopts an architecture that combines finite state machines and memory networks to maintain the context information of multi-turn dialogues and record the current stage of the credit business, the user information fields that have been collected, and the fields that need to be supplemented. The security sandbox is an independent execution environment built on TrustZone technology and virtualization isolation technology. It is used to process sensitive user data on the client side and ensure that sensitive data does not leave the sandbox. The local information collector, after obtaining explicit authorization from the user, uses the annotation API interface provided by the operating system to collect auxiliary information from legitimate information sources on the mobile device for automatic data filling.

[0008] Preferably, the lightweight large language model fine-tuned in the credit business domain specifically refers to: LoRA low-rank adaptation technique and QLoRA quantized low-rank adaptation technique are used to inject trainable low-rank matrices into the large language model; The fine-tuning data includes credit business dialogue data, financial product knowledge base, risk control rule texts, and frequently asked user questions; The fine-tuning objectives include intent recognition accuracy, entity extraction accuracy, and dialogue coherence score. The lightweight large language model, fine-tuned for the credit business, is deployed on mobile devices after being quantized with INT4 or INT8.

[0009] Preferably, the intent recognition and task orchestration module includes an intent classifier, a task orchestration engine, and an execution status monitor; The intent classifier is based on the Few-Shot learning capability of the large language model and performs accurate classification through preset credit business intent examples. The preset credit business intent examples include credit consultation, loan application, credit limit inquiry, repayment operation, bill inquiry, early repayment, information supplementation, and identity verification. The Few-Shot learning capability of the large language model is used for intent classification. Intent classification specifically includes: Build a library of credit business intent examples, with 3-5 labeled examples for each intent category; Concatenate the user's current input with the historical dialogue context to form a complete query sample; A dynamic example selection strategy is adopted, which retrieves the K most similar examples to the current query from the example library based on cosine similarity as contextual hints; The concatenated prompt is input into the large language model, which outputs the intent category label and confidence score. When the confidence score is lower than the preset threshold, a clarification dialogue is triggered to request further information from the user to eliminate intent ambiguity. The task orchestration engine automatically generates a corresponding task execution chain based on a directed acyclic graph according to the identified intent. The task execution chain includes node definition, node dependency relationship and node execution condition. The execution status monitor uses a heartbeat detection mechanism to track the execution status of each task node in real time. When a node times out or returns an exception code, it automatically triggers a retry mechanism. After the number of retries reaches a threshold, the user is asked to provide feedback on the exception and request further instructions.

[0010] Preferably, the intelligent information acquisition module includes an information source adapter, an information extraction engine, an information verifier, and an information fusion unit; The information source adapter connects to different types of mobile terminal information sources through the adapter mode. The information sources include ID card photo storage, address book, APP installation list, location information service, SMS inbox and photo album metadata. The information extraction engine employs multimodal fusion technology to perform corresponding extraction methods on different types of raw data: The algorithm uses a CRNN-based OCR text recognition algorithm to extract name, ID number, and address information from ID card photos; it uses social network analysis methods to extract the number of contacts and tightness centrality features from the address book; it uses a text classification model to extract consumption preference tags from the app installation list; and it uses a clustering algorithm to identify permanent address and work address from location information. The information validator uses a cross-validation mechanism to cross-validate the extracted information and detect the consistency of the data. The information fusion unit uses a confidence-based weighted fusion method to integrate and deduplicate information from different channels, generating complete user profile data.

[0011] Preferably, the CRNN-based OCR text recognition algorithm for extracting ID card information specifically includes: Preprocessing of ID card photos includes image correction, illumination normalization, and resolution standardization; Visual feature maps of images are extracted using convolutional neural networks; A bidirectional long short-term memory network is used to perform sequence modeling on the visual feature map, and the character probability distribution is output. End-to-end training is performed using a connection-time classification loss function, eliminating the need for character-level annotation; A dictionary-constrained decoding strategy is used to decode the probability distribution into structured text fields, including name, ID number, and address information.

[0012] Preferably, the multidimensional features identified by the AI ​​risk control model include user profile features, behavioral sequence features, social network features, and device environment features; The hard rules include anti-money laundering rules, identity verification rules, and geographical restriction rules; The edge-cloud collaborative risk control module also includes a result feedback unit, which uses an AI assistant to provide real-time feedback on risk control progress and results in natural language. The fusion decision-maker employs weighted voting or Stacking ensemble learning methods for weighted fusion, specifically including: The scoring card model outputs a credit score, which is mapped to a default probability; the AI ​​risk control model outputs a default probability value; and the rule engine outputs a binary decision result and the triggered rule code, wherein the binary decision result includes pass and reject. The Stacking ensemble learning method is adopted, with the output of the scorecard model, the output of the AI ​​risk control model, and the trigger features of the rule engine as meta-features input to the meta-classifier. The meta-classifier uses logistic regression and gradient boosting tree to output the final default probability and decision suggestions.

[0013] Preferably, the data interaction steps between the edge-side pre-evaluation unit and the cloud-based fine-evaluation unit include: After completing the pre-assessment, the edge-side pre-assessment unit generates a preliminary risk assessment report, including risk level labels and a list of abnormal features; The collected raw user data is anonymized: identity fields are hashed and salted, numeric fields are binned and encoded, and text fields are replaced with embedded vectors. The de-identified feature data, preliminary risk assessment report, and device fingerprint information are encrypted and uploaded to the cloud. After receiving the data, the cloud-based fine evaluation unit performs data integrity and format verification before executing the fine evaluation process.

[0014] Preferably, the intelligent post-loan management module includes an intelligent repayment reminder submodule, a one-click repayment execution submodule, and a financial health analysis submodule; The intelligent repayment reminder submodule generates personalized repayment reminder messages based on user profiles and historical repayment behavior through a large language model, and uses a reinforcement learning algorithm to determine the best time to push the message. The one-click repayment execution submodule accepts the user's voice repayment command and automatically executes the entire process of account verification, amount confirmation, payment channel selection, and payment execution. The financial health analysis submodule regularly analyzes users' income and expenditure, uses a time series forecasting model to predict users' future repayment ability, and when it predicts that users may have difficulty repaying, it proactively generates and pushes installment adjustment and deferred repayment suggestions.

[0015] Preferably, the intelligent repayment reminder submodule uses a reinforcement learning algorithm to determine the optimal push timing, specifically including: Define the state space, including the current time, the user's historical repayment time distribution, the user's current APP activity status, the user's recent income and expenditure trends, and the number of days until the repayment date; Define the action space, including tenderloin push, push delayed by N hours, and no push on the same day; Define a reward function, including a positive reward for users who repay on time, a negative reward for users who repay late, and a negative reward for users who turn off reminders; A deep Q-network and a policy gradient algorithm are used to train the push timing decision model; At the daily decision-making moment, the current state is input into the trained model, and the optimal push action is output.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention enables natural language interaction through a large language model, simplifying the multi-step manual operation of traditional credit business into a one-stop dialogue service. Users do not need to learn the APP operation process; they can complete all credit business by simply expressing their needs in everyday language, which significantly reduces the operation threshold and improves the user experience. This invention significantly reduces the amount of information manually entered by users through the dual-channel strategy of "active collection + dialogue completion" of the intelligent information collection module, and improves process efficiency by combining the automated process arrangement capabilities of the intent recognition and task arrangement modules. This invention adopts an edge-cloud collaborative risk control architecture to achieve risk control in advance. The edge-side pre-assessment can complete the initial risk screening in milliseconds, and the cloud-based fine assessment integrates the scoring card model, AI model and rule engine to conduct multi-dimensional evaluation. The overall risk control accuracy and efficiency are better than the traditional single back-end risk control mode. At the same time, the AI ​​assistant can provide users with real-time feedback on risk control progress, improving the transparency of the approval process. Attached Figure Description

[0017] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a structural diagram of the mobile AI assistant module of the present invention; Figure 3 This is a structural diagram of the intent recognition and task orchestration module of the present invention; Figure 4 This is a structural diagram of the intelligent information acquisition module of the present invention; Figure 5 This is a structural diagram of the edge-cloud collaborative risk control module of the present invention; Figure 6 This is a structural diagram of the intelligent post-loan management module of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below. Example 1

[0019] This embodiment provides an intelligent AI credit assistant system based on a large language model, such as... Figure 1 As shown, the entire system adopts an edge-cloud collaborative architecture, including a mobile AI assistant module, an intent recognition and task orchestration module, an intelligent information collection module, an edge-cloud collaborative risk control module, and an intelligent post-loan management module. Users interact with the mobile AI assistant module via natural language. This module incorporates a lightweight, large-scale language model fine-tuned for the credit industry, supporting both voice and text input. User input is parsed by the natural language understanding engine and then transmitted to the intent recognition and task orchestration module for intent recognition and task orchestration. During task execution, the intelligent information collection module acquires user information through a dual-channel strategy of proactive collection and dialogue completion. The edge-cloud collaborative risk control module performs two-stage risk control: edge-side pre-assessment and cloud-based fine-assessment. After loan disbursement, the intelligent post-loan management module is responsible for full lifecycle management, transforming traditional multi-step manual operations into a one-stop intelligent dialogue service. Specific modules include: The mobile AI assistant module is deployed on the user's mobile terminal and serves as the only entry point for the user to interact with the credit system. It has a built-in lightweight large language model that has been fine-tuned in the field of credit business and supports both voice and text input methods. Users can interact with the mobile AI assistant module through natural language. Users can express their credit needs through natural language, such as "I want to borrow 50,000 yuan for short-term use" or "Please help me see how much I owe this month". The lightweight large language model, fine-tuned for the credit business area, is as follows: LoRA low-rank adaptation technique and QLoRA quantized low-rank adaptation technique are used to inject trainable low-rank matrices into the large language model; The fine-tuning data includes credit business dialogue data, financial product knowledge base, risk control rule texts, and frequently asked user questions; The fine-tuning objectives include intent recognition accuracy, entity extraction accuracy, and dialogue coherence score. The lightweight large language model, fine-tuned for the credit business, is deployed on mobile devices after being quantized with INT4 or INT8. like Figure 2 As shown, the mobile AI assistant module includes a natural language understanding engine, a dialogue state manager, a security sandbox, and a local information collector: The natural language understanding engine is used to parse the semantics of user input. It adopts an encoder-decoder model based on the Transformer architecture to perform word segmentation, named entity recognition and semantic role labeling on the user's natural language input, and extract credit-related entity information (such as amount, term, product type, etc.). The dialogue state manager adopts an architecture that combines finite state machines and memory networks to maintain the context information of multi-turn dialogues, record the current stage of the credit business, the user information fields that have been collected, and the fields that need to be supplemented, so as to ensure that key information is not lost in complex credit processing procedures. The security sandbox is an independent execution environment built on TrustZone technology and virtualization isolation technology. It is used to process sensitive user data on the client side, ensuring the security of user privacy data when it is processed on the client side, and sensitive data does not leave the sandbox. The local information collector, after obtaining explicit authorization from the user, uses the annotation API interface provided by the operating system to collect auxiliary information from legitimate information sources on the mobile device for automatic data filling.

[0020] The intent recognition and task orchestration module is the "brain" of the entire system. Deployed on a cloud server, it transforms the user's natural language requirements into a sequence of executable business tasks. like Figure 3 As shown, the intent recognition and task orchestration module includes an intent classifier, a task orchestration engine, and an execution status monitor; The intent classifier is based on the Few-Shot learning capability of the large language model. It accurately classifies credit business intent examples through preset examples, including credit consultation, loan application, credit limit inquiry, repayment operation, bill inquiry, early repayment, information supplementation, and identity verification. The Few-Shot learning ability of large language models is used for intent classification. Intent classification specifically includes: Build a library of credit business intent examples, with 3-5 labeled examples for each intent category; Concatenate the user's current input with the historical dialogue context to form a complete query sample; A dynamic example selection strategy is adopted, which retrieves the K most similar examples to the current query from the example library based on cosine similarity as contextual hints; The concatenated prompt is input into the large language model, which outputs the intent category label and confidence score. When the confidence score is lower than the preset threshold, a clarification dialogue is triggered to request further information from the user to eliminate intent ambiguity. Based on the identified intent, the task orchestration engine automatically generates the corresponding task execution chain using a directed acyclic graph (DAG). The task execution chain includes node definitions, node dependencies, and node execution conditions. It retrieves the corresponding node sequence based on the intent, parses the data dependencies between nodes, constructs the DAG, performs topological sorting, and generates a linear execution sequence. During execution, nodes are dynamically skipped or inserted based on the evaluation results of the precondition expressions. For example, when the user's intent is a loan application, the generated task execution chain is as follows: identity verification node → data collection node → product matching node → risk control assessment node → credit limit approval node → contract signing node. The execution status monitor uses a heartbeat detection mechanism to track the execution status of each task node in real time. When a node times out or returns an exception code, it automatically triggers a retry mechanism. After the number of retries reaches a threshold, the user is asked to provide feedback on the exception information and request further instructions.

[0021] The intelligent information collection module adopts a dual-channel strategy of proactive collection and dialogue completion to obtain user information, which solves the problem of cumbersome data filling in traditional credit applications; like Figure 4 As shown, the intelligent information acquisition module includes an information source adapter, an information extraction engine, an information verifier, and an information fusion unit. The first channel is the active data collection channel. After obtaining explicit authorization from the user, the AI ​​assistant extracts usable information from legitimate information sources on the user's mobile phone. The information source adapter connects to different types of mobile information sources through the adapter mode. The information sources include ID card photo storage, address book, APP installation list, location information service, SMS inbox, and photo album metadata. The information extraction engine employs multimodal fusion technology to perform corresponding extraction methods on different types of raw data: for ID card photos, a CRNN-based OCR text recognition algorithm is used to extract name, ID number, and address information; for the address book, a social network analysis method is used to extract the number of contacts and closeness centrality features; for the APP installation list, a text classification model is used to extract consumption preference tags; and for location information, a clustering algorithm is used to identify permanent address and work address. The OCR text recognition algorithm based on CRNN for extracting ID card information specifically includes: Preprocessing of ID card photos includes image correction, illumination normalization, and resolution standardization; Visual feature maps of images are extracted using convolutional neural networks; A bidirectional long short-term memory network is used to perform sequence modeling on the visual feature map, and the character probability distribution is output. End-to-end training is performed using a connection-time classification loss function, eliminating the need for character-level annotation; A dictionary-constrained decoding strategy is used to decode the probability distribution into structured text fields, including name, ID number, and address information; The second channel is the dialogue completion channel. For information that cannot be obtained through active collection or that requires user confirmation, the AI ​​assistant guides the user to complete it through natural dialogue. The information verifier uses a cross-validation mechanism to cross-validate the extracted information and detect the consistency of the data: it calculates the text similarity between the ID card address recognized by OCR and the permanent address recognized by location clustering (such as using Jaccard similarity or cosine similarity), and marks the information as unverified when the similarity is lower than a preset threshold. The information fusion engine uses a confidence-based weighted fusion method to integrate and deduplicate information from different channels, generating complete user profile data.

[0022] The edge-cloud collaborative risk control module breaks down risk control assessment into two stages: edge-side pre-assessment and cloud-based fine assessment. This enables proactive risk control and real-time feedback, and achieves deep collaboration between mobile AI analysis and remote risk control systems. Traditional risk control systems only run on remote servers, and users need to wait for backend approval after submitting an application. like Figure 5 As shown, the edge-cloud collaborative risk control module includes an edge-side pre-assessment unit and a cloud-based fine-assessment unit; The edge-side pre-assessment unit includes a lightweight risk control model deployed on the mobile device. It performs preliminary risk screening based on locally collected user behavior data. It adopts a lightweight model based on gradient boosting decision trees or a neural network model compressed by knowledge distillation. The user behavior data includes: operation habit sequences (such as click hotspot distribution, page dwell time), form filling speed, information modification frequency, device fingerprint information (such as device model, operating system version, screen resolution), and input trajectory features (such as touch pressure, swipe speed). The edge-side pre-assessment is executed locally with a latency controlled within 50 milliseconds. It can quickly identify obvious fraud characteristics (such as device abnormality, information forgery, robot operation, etc.) and provide preliminary feedback before the user submits the application. The cloud-based precision assessment unit is deployed on a remote server and includes a scoring card model, an AI risk control model, a rules engine, and a fusion decision-maker. The credit scoring model uses logistic regression and linear regression algorithms to score credit based on factors such as user credit history, income level, debt ratio, and job stability. AI risk control models use deep neural networks and graph neural networks to identify potential risk patterns from multi-dimensional features, including user profile features, behavioral sequence features, social network features, and device environment features. The rules engine performs hard rule checks for regulatory compliance, including anti-money laundering rules, identity verification rules, and geographical restriction rules; The fusion decision-maker uses weighted voting and stacking ensemble learning methods to weightedly fuse the evaluation results of the scorecard model, AI risk control model, and rule engine to generate the final risk control decision result. The fusion decision-maker employs weighted voting or Stacking ensemble learning methods for weighted fusion, specifically including: The scoring card model outputs a credit score, which is mapped to a default probability; the AI ​​risk control model outputs a default probability value; and the rule engine outputs a binary decision result and the triggered rule code, wherein the binary decision result includes pass and reject. The Stacking ensemble learning method is adopted, with the output of the scorecard model, the output of the AI ​​risk control model, and the trigger features of the rule engine as meta-features input to the meta-classifier. The meta-classifier uses logistic regression and gradient boosting tree to output the final default probability and decision suggestions. The results feedback unit uses an AI assistant to provide real-time feedback on risk control progress and results in natural language. The data interaction steps between the edge-side pre-evaluation unit and the cloud-based fine-evaluation unit include: After completing the pre-assessment, the edge-side pre-assessment unit generates a preliminary risk assessment report, including risk level labels and a list of abnormal features; The collected raw user data is anonymized: identity fields are hashed and salted, numeric fields are binned and encoded, and text fields are replaced with embedded vectors. The de-identified feature data, preliminary risk assessment report, and device fingerprint information are encrypted and uploaded to the cloud. After receiving the data, the cloud-based precision evaluation unit performs data integrity and format verification before executing the precision evaluation process. The intelligent post-loan management module is used to perform proactive post-loan management, and is responsible for the full life cycle management after the loan is issued. It adopts an AI-driven proactive post-loan management strategy. like Figure 6 As shown, the intelligent post-loan management module includes an intelligent repayment reminder sub-module, a one-click repayment execution sub-module, and a financial health analysis sub-module; The intelligent repayment reminder submodule generates personalized repayment reminder messages based on user profiles and historical repayment behavior through a large language model, and uses reinforcement learning algorithms to determine the best time to push the message. The intelligent repayment reminder submodule uses reinforcement learning algorithms to determine the optimal timing for sending notifications, specifically including: Define the state space, including the current time, the user's historical repayment time distribution, the user's current APP activity status, the user's recent income and expenditure trends, and the number of days until the repayment date; Define the action space, including tenderloin push, push delayed by N hours, and no push on the same day; Define a reward function, including a positive reward for users who repay on time, a negative reward for users who repay late, and a negative reward for users who turn off reminders; A deep Q-network and a policy gradient algorithm are used to train the push timing decision model; At the daily decision-making moment, the current state is input into the trained model, and the optimal push action is output. The one-click repayment execution submodule accepts the user's voice repayment command and automatically executes the entire process of account verification, amount confirmation, payment channel selection, and payment execution. The financial health analysis submodule regularly analyzes users' income and expenditure, uses time series forecasting models to predict users' future repayment ability, and proactively generates and pushes installment adjustment and deferred repayment suggestions when it predicts that users may have repayment difficulties.

[0023] The system provided in this embodiment can be widely applied in the following fields: (1) Consumer finance sector: Personal consumer loan business of institutions such as banks and consumer finance companies, providing users with convenient loan application and management services through AI assistants; (2) Micro and small enterprise credit field: Provide intelligent business loan application services for micro and small business owners. The AI ​​assistant automatically collects business data (such as business license, tax records, and bank statements) and completes risk assessment; (3) Credit card business area: Intelligent management of credit card application, credit limit adjustment, bill installment and other businesses, replacing traditional form operations with dialogue interaction; (4) Internet finance platforms: The intelligent customer service of various online lending platforms has been upgraded from simple Q&A to full-process business processing capabilities; (5) Bank mobile banking APP: Embedded as a smart assistant module of mobile banking, it enhances the service capabilities and user experience of mobile banking in credit business.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart AI credit assistant system based on a large language model, characterized in that, It includes a mobile AI assistant module, an intent recognition and task orchestration module, an intelligent information collection module, an edge-cloud collaborative risk control module, and an intelligent post-loan management module; The mobile AI assistant module is deployed on the user's mobile terminal and has a built-in lightweight large language model that has been fine-tuned in the field of credit business. Users interact with the mobile AI assistant module through natural language. The intent recognition and task orchestration module is deployed on a cloud server and transforms the user's natural language requirements into a sequence of executable business tasks. The intelligent information collection module adopts a dual-channel strategy of active collection and dialogue completion to obtain user information; The edge-cloud collaborative risk control module divides the risk control assessment into two stages: edge-side pre-assessment and cloud-based fine assessment, realizing risk control in advance and real-time feedback. The edge-cloud collaborative risk control module includes an edge-side pre-assessment unit and a cloud-based fine assessment unit. The edge-side pre-assessment unit includes a lightweight risk control model deployed on a mobile device. It adopts a lightweight model based on gradient boosting decision trees or a neural network model compressed by knowledge distillation, and performs preliminary risk screening based on locally collected user behavior data. The cloud-based assessment unit, deployed on a remote server, includes a credit scoring model, an AI risk control model, a rules engine, and a fusion decision-maker. The credit scoring model uses logistic regression and linear regression algorithms to score credit based on user credit history, income level, debt ratio, and job stability. The AI ​​risk control model uses deep neural networks and graph neural networks to identify potential risk patterns from multi-dimensional features. The rules engine performs mandatory regulatory compliance checks. The fusion decision-maker uses weighted voting and stacking ensemble learning methods to weightedly fuse the assessment results from the credit scoring model, AI risk control model, and rules engine to generate the final risk control decision. The intelligent post-loan management module is used to perform proactive post-loan management.

2. The intelligent AI credit assistant system based on a large language model according to claim 1, characterized in that, The mobile AI assistant module serves as the sole entry point for users to interact with the credit system, supporting both voice and text input methods. The mobile AI assistant module includes: The natural language understanding engine is used to parse the semantics of user input. It adopts an encoder-decoder model based on the Transformer architecture to perform word segmentation, named entity recognition, and semantic role labeling on the user's natural language input, and extract credit-related entity information. The dialogue state manager adopts an architecture that combines finite state machines and memory networks to maintain the context information of multi-turn dialogues and record the current stage of the credit business, the user information fields that have been collected, and the fields that need to be supplemented. The security sandbox is an independent execution environment built on TrustZone technology and virtualization isolation technology. It is used to process sensitive user data on the client side and ensure that sensitive data does not leave the sandbox. The local information collector, after obtaining explicit authorization from the user, uses the annotation API interface provided by the operating system to collect auxiliary information from legitimate information sources on the mobile device for automatic data filling.

3. The intelligent AI credit assistant system based on a large language model according to claim 2, characterized in that, The lightweight large language model fine-tuned in the credit business area is specifically as follows: LoRA low-rank adaptation technique and QLoRA quantized low-rank adaptation technique are used to inject trainable low-rank matrices into the large language model; The fine-tuning data includes credit business dialogue data, financial product knowledge base, risk control rule texts, and frequently asked user questions; The fine-tuning objectives include intent recognition accuracy, entity extraction accuracy, and dialogue coherence score. The lightweight large language model, fine-tuned for the credit business, is deployed on mobile devices after being quantized with INT4 or INT8.

4. The intelligent AI credit assistant system based on a large language model according to claim 1, characterized in that, The intent recognition and task orchestration module includes an intent classifier, a task orchestration engine, and an execution status monitor. The intent classifier is based on the Few-Shot learning capability of the large language model and performs accurate classification through preset credit business intent examples. The preset credit business intent examples include credit consultation, loan application, credit limit inquiry, repayment operation, bill inquiry, early repayment, information supplementation, and identity verification. The Few-Shot learning capability of the large language model is used for intent classification. Intent classification specifically includes: Build a library of credit business intent examples, with 3-5 labeled examples for each intent category; Concatenate the user's current input with the historical dialogue context to form a complete query sample; A dynamic example selection strategy is adopted, which retrieves the K most similar examples to the current query from the example library based on cosine similarity as contextual hints; The concatenated prompt is input into the large language model, which outputs the intent category label and confidence score. When the confidence score is lower than the preset threshold, a clarification dialogue is triggered to request further information from the user to eliminate intent ambiguity. The task orchestration engine automatically generates a corresponding task execution chain based on a directed acyclic graph according to the identified intent. The task execution chain includes node definition, node dependency relationship and node execution condition. The execution status monitor uses a heartbeat detection mechanism to track the execution status of each task node in real time. When a node times out or returns an exception code, it automatically triggers a retry mechanism. After the number of retries reaches a threshold, the user is asked to provide feedback on the exception and request further instructions.

5. The intelligent AI credit assistant system based on a large language model according to claim 1, characterized in that, The intelligent information acquisition module includes an information source adapter, an information extraction engine, an information verifier, and an information fusion unit; The information source adapter connects to different types of mobile terminal information sources through the adapter mode. The information sources include ID card photo storage, address book, APP installation list, location information service, SMS inbox and photo album metadata. The information extraction engine employs multimodal fusion technology to perform corresponding extraction methods on different types of raw data: The ID card photo was processed using a CRNN-based OCR text recognition algorithm to extract the name, ID number, and address information; the address book was processed using social network analysis methods to extract the number of contacts and the closeness centrality feature. A text classification model is used to extract consumer preference tags from the app installation list; a clustering algorithm is used to identify permanent and work addresses from location information. The information validator uses a cross-validation mechanism to cross-validate the extracted information and detect the consistency of the data; The information fusion unit uses a confidence-based weighted fusion method to integrate and deduplicate information from different channels, generating complete user profile data.

6. The intelligent AI credit assistant system based on a large language model according to claim 5, characterized in that, The OCR text recognition algorithm based on CRNN for extracting ID card information specifically includes: Preprocessing of ID card photos includes image correction, illumination normalization, and resolution standardization; Visual feature maps of images are extracted using convolutional neural networks; A bidirectional long short-term memory network is used to perform sequence modeling on the visual feature map, and the character probability distribution is output. End-to-end training is performed using a connection-time classification loss function, eliminating the need for character-level annotation; A dictionary-constrained decoding strategy is used to decode the probability distribution into structured text fields, including name, ID number, and address information.

7. The intelligent AI credit assistant system based on a large language model according to claim 1, characterized in that, The multidimensional features identified by the AI ​​risk control model include user profile features, behavioral sequence features, social network features, and device environment features; The hard rules include anti-money laundering rules, identity verification rules, and geographical restriction rules; The edge-cloud collaborative risk control module also includes a result feedback unit, which uses an AI assistant to provide real-time feedback on risk control progress and results in natural language. The fusion decision-maker employs weighted voting or Stacking ensemble learning methods for weighted fusion, specifically including: The scoring card model outputs a credit score, which is mapped to a default probability; the AI ​​risk control model outputs a default probability value; and the rule engine outputs a binary decision result and the triggered rule code, wherein the binary decision result includes pass and reject. The Stacking ensemble learning method is adopted, with the output of the scorecard model, the output of the AI ​​risk control model, and the trigger features of the rule engine as meta-features input to the meta-classifier. The meta-classifier uses logistic regression and gradient boosting tree to output the final default probability and decision suggestions.

8. The intelligent AI credit assistant system based on a large language model according to claim 7, characterized in that, The data interaction steps between the edge-side pre-evaluation unit and the cloud-based fine-evaluation unit include: After completing the pre-assessment, the edge-side pre-assessment unit generates a preliminary risk assessment report, including risk level labels and a list of abnormal features; The collected raw user data is anonymized: identity fields are hashed and salted, numeric fields are binned and encoded, and text fields are replaced with embedded vectors. The de-identified feature data, preliminary risk assessment report, and device fingerprint information are encrypted and uploaded to the cloud. After receiving the data, the cloud-based fine evaluation unit performs data integrity and format verification before executing the fine evaluation process.

9. The intelligent AI credit assistant system based on a large language model according to claim 1, characterized in that, The intelligent post-loan management module includes an intelligent repayment reminder sub-module, a one-click repayment execution sub-module, and a financial health analysis sub-module. The intelligent repayment reminder submodule generates personalized repayment reminder messages based on user profiles and historical repayment behavior through a large language model, and uses a reinforcement learning algorithm to determine the best time to push the message. The one-click repayment execution submodule accepts the user's voice repayment command and automatically executes the entire process of account verification, amount confirmation, payment channel selection, and payment execution. The financial health analysis submodule regularly analyzes users' income and expenditure, uses a time series forecasting model to predict users' future repayment ability, and when it predicts that users may have difficulty repaying, it proactively generates and pushes installment adjustment and deferred repayment suggestions.

10. The intelligent AI credit assistant system based on a large language model according to claim 9, characterized in that, The intelligent repayment reminder submodule uses a reinforcement learning algorithm to determine the optimal timing for sending notifications, specifically including: Define the state space, including the current time, the user's historical repayment time distribution, the user's current APP activity status, the user's recent income and expenditure trends, and the number of days until the repayment date; Define the action space, including tenderloin push, push delayed by N hours, and no push on the same day; Define a reward function, including a positive reward for users who repay on time, a negative reward for users who repay late, and a negative reward for users who turn off reminders; A deep Q-network and a policy gradient algorithm are used to train the push timing decision model; At the daily decision-making moment, the current state is input into the trained model, and the optimal push action is output.

Images