Generative artificial intelligence scientific and technological innovation incubation digital broker system and interaction method

Abstract

The present application relates to the field of information technology and artificial intelligence technology, in particular to a digital broker system for scientific and technological innovation incubation based on generative artificial intelligence, comprising: a user management module for supporting user registration, login and user role management; a user portrait analysis module connected to the user management module for collecting user portrait data through user registration information, usage behavior, access records and operation habits; the present application generates a user portrait by collecting multi-dimensional data such as user registration information, usage behavior, access records and operation habits through the user portrait analysis module, and automatically generates a customized broker digital image matched with the user portrait by calling a generative AI model from a personalized digital broker generation module, thereby realizing deep personalized service for one user and one image, and solving the technical problem that traditional platforms cannot provide differentiated service experience according to the personalized characteristics of user industry background, entrepreneurial stage and technical field.

Description

Technical Field

[0001] This invention relates to the fields of information technology and artificial intelligence technology, and in particular to a digital brokerage system and interaction method for incubating science and technology innovation based on generative artificial intelligence. Background Technology

[0002] With the deepening of technology transfer and innovation and entrepreneurship services, science and technology incubation service platforms face challenges such as insufficient personalized services, low efficiency in responding to professional consultations, and a limited user experience. Traditional incubation service platforms mainly rely on human experts for consultation and services, which has the following shortcomings: First, the service model is too simplistic and it is difficult to provide personalized services. Traditional incubation service platforms usually use a unified service interface and standardized information push methods, which cannot provide differentiated service experiences based on the user's industry background, entrepreneurial stage, technical field and other personalized characteristics.

[0003] Second, professional service resources are scattered and response efficiency is low. Startup incubation involves multiple professional fields such as law, finance and taxation, intellectual property, and business management. Traditional platforms find it difficult to provide users with comprehensive professional services across fields through a single entry point, and users often need to switch between different channels.

[0004] Third, the interaction methods are limited, and the user experience is restricted. Existing platforms mainly rely on text interaction and lack multimodal interaction methods such as voice, 3D visualization, and data visualization, making it difficult to meet the diverse interaction needs of users.

[0005] Fourth, the matching efficiency of supply and demand for scientific and technological achievements is low. The traditional model relies on human experience and personal recommendations to match achievements with needs, resulting in low matching accuracy and long cycles.

[0006] In recent years, AI digital human technology has been applied in customer service, and virtual character generation technology has also been explored. However, these technologies are mainly geared towards general customer service or entertainment scenarios, and a digital brokerage system specifically designed for the vertical field of science and technology innovation incubation, integrating generative AI large models, multi-role virtual experts, and 3D visualization interaction, has yet to emerge. Therefore, a generative AI-based digital brokerage system and interaction methods for science and technology innovation incubation are needed to address these issues. Summary of the Invention

[0007] The purpose of this invention is to solve the problems mentioned in the background section.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A generative AI-based digital brokerage system for science and technology innovation incubation includes: The user management module supports user registration, login, and user role management. The user profile analysis module, connected to the user management module, is used to collect user profile data through user registration information, usage behavior, access records and operation habits, use big data analysis technology to extract user interests, usage habits and needs preferences to generate user profiles, and dynamically update user preferences and interest tags according to changes in user behavior. The personalized digital agent generation module is connected to the user profile analysis module. It is used to automatically generate a customized digital agent image based on the user profile through a generative AI model, so as to achieve one image per user. The underlying framework for AI services is used to provide underlying support for model training, model management, and model application. Generative AI large language model, deployed on the underlying framework of the AI ​​service, includes at least a startup incubation knowledge base model, a corporate legal knowledge base model, a corporate financial knowledge base model, a corporate tax knowledge base model, and a planning text model; The multi-role virtual expert service module connects the generative artificial intelligence big language model and the personalized digital broker generation module. Based on the user profile and the big language model, it provides users with virtual entrepreneurial mentor services, virtual lawyer services, virtual financial and tax expert services, virtual intellectual property consultant services, and virtual planning text writing expert services. A multimodal interactive display platform, connecting the personalized digital agent generation module and the multi-role virtual expert service module, is used to realize multimodal interaction between users and digital agents through at least two of the following methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction. The intelligent recommendation module connects the user profile analysis module and the multi-role virtual expert service module. It is used to intelligently recommend content and services that users are interested in based on the user profile, and dynamically optimize the recommendation algorithm according to the user's real-time behavior and feedback. The data management module connects the user profile analysis module and the underlying AI service framework. It is used to collect data through multiple channels, clean, store and analyze the collected data to uncover user needs and behavioral patterns.

[0009] Preferably, the underlying framework of the AI ​​service includes: a data acquisition module for acquiring massive amounts of data from multiple data sources; A network configuration module used to provide high-bandwidth, low-latency network connectivity and load balancing; A service deployment module for enabling rapid deployment and elastic scaling of AI models through containerization technology; A service monitoring module for real-time monitoring and logging of all stages of model training and application; A security and compliance module used to protect data privacy through data encryption and access control; A performance optimization module for improving model inference speed through hardware accelerators.

[0010] Preferably, in the generative artificial intelligence large language model: The startup incubation knowledge base model covers startup theory, business models, and market analysis knowledge; The corporate legal knowledge base model covers legal regulations, compliance requirements, and contract management knowledge. The enterprise financial knowledge base model covers knowledge of financial management, budgeting, and capital operation. The enterprise tax knowledge base model provides information on tax planning, tax declaration, and tax compliance. The proposed text model is used for the generation and optimization of business plans, project reports, and marketing copy.

[0011] Preferably, in the multi-role virtual expert service module: The virtual entrepreneurship mentor service is used to provide assistance with entrepreneurship direction selection, business model design, market research, competitive analysis, success case analysis, and resource matching. The virtual lawyer service is used to provide legal advice and legal risk assessment on company establishment, contract review, labor law, and intellectual property law. The virtual financial and tax expert service is used to provide financial planning, budget management, fundraising, tax planning, and financial statement analysis. The virtual intellectual property advisory service is used to provide guidance on patent applications, consultation on trademark registration, and assessment of intellectual property infringement risks. The virtual project writing expert service provides guidance on writing business plans, project proposals, and marketing plans.

[0012] Preferably, the multimodal interactive display platform includes: A 3D underlying rendering engine based on WebGL technology and using a B / S architecture; An interactive data presentation layer used for visualizing data in chart form; A 3D model interactive display layer used to support 3D model loading and rendering operations; A voice interaction presentation layer used to support voice input recognition and voice feedback; And an image interaction display layer to support image display and interactive operations.

[0013] Preferably, the 3D underlying rendering engine supports loading 3D models in OBJ, FBX, GLTF and STL formats, implements PBR physically based rendering materials, and supports multi-light source lighting effects as well as model rotation, scaling, translation, selection and highlighting. The data interaction presentation layer integrates ECharts or D3.js chart libraries, and uses WebSocket or Server-SentEvents to achieve real-time data push and dynamic chart updates, and provides data filtering, sorting, filtering and data export functions.

[0014] Preferably, the voice interaction presentation layer integrates a speech recognition engine and a speech synthesis engine, and uses natural language processing technology to realize multi-turn dialogue, contextual understanding, and automatic detection and switching of multilingual input and output; The image interactive display layer supports loading images in JPEG, PNG and SVG formats, provides gallery mode, slideshow mode and single image mode display modes, and implements image zooming, dragging, cropping, rotating, annotation and sharing functions.

[0015] Preferably, it also includes a generative artificial intelligence image model, which is a diffusion model, connected to the multi-role virtual expert service module, and used for chart generation, effect image generation, and PPT generation to assist the output results of the multi-role virtual expert service module.

[0016] Preferably, it also includes a dataset expansion module, which is connected to the underlying framework of the AI ​​service and is used to perform manual data iteration through data construction, data collection, prompt word labeling, data classification and multi-source data feeding strategies to achieve dataset expansion and optimization.

[0017] The interactive method for digital brokerage incubation of science and technology innovation based on generative artificial intelligence includes the following steps: Step 1: Obtain user data through the user management module, and generate and dynamically update user profiles through the user profile analysis module; Step 2: Based on the user profile, the personalized digital agent generation module calls the generative AI model to automatically generate a customized digital agent image. Step 3: Receive user interaction requests through the multi-role virtual expert service module, identify user intent based on the generative artificial intelligence large language model, and call the corresponding professional model from the startup incubation knowledge base model, corporate legal knowledge base model, corporate financial knowledge base model, corporate tax knowledge base model and planning text model to match and activate the corresponding virtual expert role. Step 4: Through the multimodal interactive display platform, the service results are output to the user using at least two of the following interaction methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction. Service content is also recommended to the user through the intelligent recommendation module. Step 5: Record user feedback and behavioral data through the data management module, update the user profile, and feed the updated data into the generative artificial intelligence large language model for iterative optimization.

[0018] The present invention has at least the following beneficial effects: 1. This invention generates user profiles by collecting multi-dimensional data such as user registration information, usage behavior, access records, and operating habits through a user profile analysis module. The personalized digital broker generation module calls a generative AI model to automatically generate a customized digital broker image that matches the user profile, realizing a deep personalized service with one image per user. This solves the technical problem that traditional platforms cannot provide differentiated service experiences based on the personalized characteristics of users such as industry background, entrepreneurial stage, and technical field.

[0019] 2. This invention integrates five professional roles—virtual entrepreneurship mentor, virtual lawyer, virtual financial and tax expert, virtual intellectual property consultant, and virtual business plan writing expert—into a single digital brokerage system. Through the multi-role virtual expert service module, the system intelligently identifies user intent based on a large language model and automatically activates the corresponding expert role. Users can obtain comprehensive professional services across different fields without switching between different platforms, thus improving service response efficiency and convenience.

[0020] 3. This invention integrates a 3D underlying rendering engine, a data interaction display layer, a 3D model interaction display layer, a voice interaction display layer, and an image interaction display layer through a multimodal interactive display platform. It enables multimodal interaction between users and digital brokers through at least two of the following methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction. This breaks through the limitations of traditional platforms that are mainly text-based, and significantly improves the naturalness and immersion of user interaction.

[0021] This invention utilizes a generative artificial intelligence large language model to perform semantic understanding and intelligent matching of user needs and scientific and technological achievements. Through an intelligent recommendation module, it accurately pushes achievements, services, and information that users are interested in based on user profiles. This AI-driven supply and demand matching replaces the traditional docking mode that relies on human experience and recommendations from acquaintances, significantly improving matching accuracy and docking efficiency, and shortening the achievement transformation cycle. Through a data management module, user feedback and behavioral data are recorded and continuously fed into the large language model for iterative optimization, forming a closed-loop mechanism of data collection, profile generation, service output, and feedback iteration, ensuring the continuous improvement of the system's service capabilities. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the overall modular architecture of the system of the present invention; Figure 2 This is a schematic diagram of the digital broker interaction method of the present invention; Figure 3 This is a schematic diagram of the personalized digital broker generation process of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0025] Reference Figure 1-3 ; Example 1:

[0026] This embodiment is applicable to scenarios where operators of science and technology innovation incubation service platforms deploy AI digital broker systems to provide technology transfer and business incubation services for universities, research institutes, startups and entrepreneurial teams. The system is deployed on the server cluster of the Western China Smart Incubation Service Platform and provides services to users in six western provinces and even the whole country via the Internet.

[0027] The science and technology innovation incubation digital brokerage system based on generative artificial intelligence in this embodiment includes the following modules: The user management module connects to the user profile analysis module. After a user completes registration and login through the user management module, the user management module transmits the user's registration information to the user profile analysis module.

[0028] The user profile analysis module is connected to the personalized digital agent generation module, the intelligent recommendation module, and the data management module. The generated user profiles are transmitted to the personalized digital agent generation module (for generating personalized agent profiles) and the intelligent recommendation module (for personalized recommendations), while the raw data and analysis results are transmitted to the data management module for storage.

[0029] The AI ​​service underlying framework is deployed at the lowest level of the system, and the generative artificial intelligence large language model is deployed on top of the AI ​​service underlying framework, relying on the support capabilities provided by the underlying framework such as data collection, network configuration, service deployment, service monitoring, security compliance, and performance optimization.

[0030] Generative AI large language model connects to multi-role virtual expert service module, providing reasoning services to multi-role virtual expert service module for startup incubation knowledge base model, corporate legal knowledge base model, corporate financial knowledge base model, corporate tax knowledge base model and planning text model.

[0031] The multi-role virtual expert service module is connected to the personalized digital agent generation module, the multimodal interactive display platform, and the intelligent recommendation module. The multi-role virtual expert service module obtains agent profile configurations from the personalized digital agent generation module, retrieves user preference information from the intelligent recommendation module, and transmits the service processing results to the multimodal interactive display platform for output.

[0032] The data management module connects to the user profiling analysis module and the underlying AI service framework to achieve unified management and flow of data.

[0033] The system works as follows: First, users complete registration and login through the user management module. The user management module then transmits the user's registration information (including industry sector, identity type, etc.) to the user profile analysis module.

[0034] Then, the user profile analysis module continuously collects user behavior, access records and operation habits data during the user's use. It uses big data analysis technology to extract user interests, usage habits and needs preferences, generate user profiles, and dynamically update user preferences and interest tags based on changes in user behavior.

[0035] Next, the personalized digital agent generation module uses a generative AI model to automatically generate a customized digital agent avatar (including appearance features and professional competence configuration) that matches the received user profile, achieving one avatar per user.

[0036] Meanwhile, when a user initiates an interaction request through the multimodal interactive platform, the multi-role virtual expert service module receives the request, identifies the user's intent based on the generative artificial intelligence large language model, and calls professional knowledge from the corresponding knowledge base model according to the identification results to match and activate the corresponding virtual expert role, such as virtual entrepreneurial mentor, virtual lawyer, virtual financial and tax expert, virtual intellectual property consultant or virtual planning text writing expert.

[0037] Then, the multimodal interactive display platform outputs service results to users through at least two of the following methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction.

[0038] Finally, the intelligent recommendation module recommends content and services that users are interested in based on user profiles, while the data management module records user feedback and behavioral data, updates user profiles, and feeds the updated data into a generative artificial intelligence large language model for iterative optimization.

[0039] All the above modules are connected through the system bus or network interface for data communication. Each module is deployed in the server cluster as a software module and interacts with the modules through HTTP, HTTPS or RPC framework. The user management module adopts a unified authentication mechanism for multiple platforms based on the OAuth2.0 protocol, and the user profile analysis module adopts a distributed big data processing framework based on Spark or Flink for real-time analysis of user behavior data.

[0040] The generative AI model called by the personalized digital broker generation module is an image generation model based on StableDiffusion or a similar architecture. It is deployed after being fine-tuned and trained for the science and technology incubation scenario. The specific implementation of the above modules can adopt conventional technical means known in the art. Those skilled in the art can implement the system based on the description in this specification.

[0041] Example 2: This embodiment is applicable to providing stable, efficient, and secure underlying infrastructure support for digital brokerage systems in large-scale AI model training and inference scenarios. The system is deployed in a hybrid cloud environment, balancing data security and computing elasticity.

[0042] The underlying framework of the AI ​​service consists of the following six sub-modules, which are interconnected via a high-speed internal network: The data acquisition module connects to external data sources, including user behavior databases, third-party API interfaces, and file storage systems. The acquired data is preprocessed and then transmitted to the data transmission channel managed by the network configuration module.

[0043] The network configuration module connects to the service deployment module, providing high-bandwidth, low-latency network connectivity for model training and inference tasks, and enabling intelligent traffic management and load balancing.

[0044] The service deployment module uses containerization technology to deploy the trained AI model to the computing node and connects to the service monitoring module, which then monitors the deployed service in real time.

[0045] The service monitoring module connects to the security and compliance module and the performance optimization module respectively. It transmits security events in the monitoring data to the security and compliance module for processing, and transmits performance bottleneck data to the performance optimization module for analysis and optimization.

[0046] Both the security and compliance module and the performance optimization module are connected to each computing node and storage node of the system to achieve security protection and performance tuning for the entire system.

[0047] The working process of the underlying framework for AI services is as follows: First, the data acquisition module quickly and accurately acquires massive amounts of data from various data sources, including user behavior databases, third-party data interfaces, and file storage systems. It supports multiple data formats (JSON, XML, CSV, Parquet, etc.) and protocols (HTTP, JDBC, MQ, etc.), and has high throughput and low latency data acquisition capabilities.

[0048] Then, the network configuration module provides high-bandwidth, low-latency network connectivity for the collected data transmission, and ensures the efficiency and stability of data transmission between various computing nodes through intelligent traffic management and load balancing strategies.

[0049] Next, the service deployment module uses automated deployment tools and containerization technologies (such as Docker and Kubernetes) to quickly deploy the trained AI model to different runtime environments, and supports distributed deployment and elastic scaling to meet service needs of different scales.

[0050] Meanwhile, the service monitoring module monitors and logs each stage of model training and application in real time, promptly identifying and resolving potential problems to ensure high availability and reliability of the service.

[0051] Then, the security and compliance module protects data privacy and model security through measures such as data encryption, access control, and authentication, ensuring that the system complies with relevant laws, regulations, and industry standards at every stage of data collection, storage, transmission, and processing.

[0052] Finally, the performance optimization module continuously analyzes and optimizes the system's performance through high-performance computing technology, optimization algorithms, and hardware accelerators (such as GPUs and TPUs), significantly improving the speed of model training and inference. At the same time, the system has good scalability and can be smoothly expanded as business needs grow.

[0053] The data acquisition module supports data source types including relational databases (MySQL, PostgreSQL), non-relational databases (MongoDB, Redis), object storage (MinIO, AWS S3), message queues (Kafka, RabbitMQ), and RESTful API data sources. Data acquisition employs a strategy combining incremental and full data collection. Incremental collection achieves real-time data synchronization by monitoring the change logs of the data source (such as MySQL's binlog). The network configuration module employs Software-Defined Networking (SDN) technology to achieve fine-grained management and dynamic scheduling of network traffic, while load balancing uses either weighted round-robin or consistent hashing algorithms. The service deployment module's containerized orchestration is based on Kubernetes, supporting HPA (HorizontalPodAutoscaler) for automatic scaling. The service monitoring module uses Prometheus to collect metric data, Grafana for visualization, and the ELK (Elasticsearch-Logstash-Kibana) stack for log collection, storage, and analysis. The security and compliance module uses AES-256 symmetric encryption combined with RSA asymmetric encryption for key exchange. Access control employs a role-based access control (RBAC) model. The performance optimization module utilizes hardware accelerators including NVIDIA GPUs (for model training) and Intel DLBoost (for inference acceleration). All the above-mentioned technical methods are conventional implementations known in the art, and those skilled in the art can implement this underlying framework based on the guidance in this specification.

[0054] Example 3: Implementation methods for generative artificial intelligence large language models and multi-role virtual expert services.

[0055] The generative AI large language model consists of five specialized sub-models, each independent of the others but sharing the underlying base model parameters: The startup incubation knowledge base model connects to the virtual startup mentor service in the multi-role virtual expert service module, providing knowledge support for startup guidance.

[0056] The corporate legal knowledge base model connects the virtual lawyer service and virtual intellectual property advisor service in the multi-role virtual expert service module, providing knowledge support for legal consultation and intellectual property services.

[0057] The enterprise financial knowledge base model and the enterprise tax knowledge base model are jointly connected to the virtual financial and tax expert service in the multi-role virtual expert service module, providing knowledge support for financial and tax services.

[0058] The planning text model connects to the virtual planning text writing expert service in the multi-role virtual expert service module, providing capability support for text generation.

[0059] The five virtual expert services in the multi-role virtual expert service module are all connected to the personalized digital broker generation module, which configures the presentation and interaction style of the corresponding expert roles according to the user profile.

[0060] Taking user interaction with the system as an example, the workflow of multi-role virtual expert services is as follows: First, users register through the user management module, filling in information such as industry sector (AI healthcare) and identity (startup founder). Based on this information and the user's subsequent usage behavior, the user profiling and analysis module generates user profiles tagged with elements such as AI healthcare industry, startup stage, and need for legal and financial services.

[0061] Then, based on the user profile, the personalized digital agent generation module calls a generative AI model to generate a professional and approachable customized digital agent avatar for the user.

[0062] Next, the user demonstrates the platform's voice interaction function through multimodal interaction by saying, "I would like to know what legal procedures are required to establish an AI healthcare company, and also help me with a financial budget." After receiving this request, the multi-role virtual expert service module calls upon the generative artificial intelligence large language model to perform intent recognition.

[0063] Next, the intent identification results include two service requests: legal advice for company establishment and financial budget preparation. Based on the identification results, the multi-role virtual expert service module simultaneously activates virtual lawyer services and virtual financial and tax expert services.

[0064] The virtual lawyer service utilizes a corporate legal knowledge base model to obtain information such as relevant laws, regulations, and compliance requirements for company establishment, and generates legal guidance solutions for the establishment of companies in the AI ​​healthcare industry, including suggestions on company type selection, registered capital, and industry access permit requirements.

[0065] The virtual financial and tax expert service utilizes enterprise financial knowledge base models and enterprise tax knowledge base models, combined with the characteristics of the AI ​​healthcare industry, to generate financial budget templates and tax planning suggestions.

[0066] Finally, the multimodal interactive platform displays the images of virtual lawyers and virtual financial and tax experts through 3D model interaction, displays financial budget tables and charts through data visualization interaction, and outputs consultation results through voice interaction, providing users with a one-stop multi-role comprehensive consultation service.

[0067] The base model for the generative AI large language model can utilize open-source large language models such as LLaMA, Qwen, and ChatGLM, with a parameter count in the tens of billions. Five specialized sub-models are obtained through fine-tuning training on domain-specific data or efficient parameter fine-tuning (such as LoRA) based on the base model. Training data for the startup incubation knowledge base model includes books on startup management, business case libraries, and market research reports; training data for the corporate legal knowledge base model includes legal databases, contract template libraries, and judicial interpretations; training data for the corporate finance knowledge base model includes financial management textbooks and listed company financial reports; training data for the corporate tax knowledge base model includes tax law provisions, tax cases, and tax payment guides; and training data for the planning text model includes business plan templates, project application examples, and marketing copy.

[0068] All training data were obtained from publicly available and compliant data sources. Intent recognition employs a few-shot prompting method based on a large language model. By designing prompt templates containing service domain classification labels, the large language model performs semantic understanding and domain classification of user input. Multi-role parallel activation involves simultaneously calling multiple sub-models via multi-threading or multi-processing. The inference results of each sub-model are integrated and output uniformly through a result fusion module. All the specific implementation methods described above are conventional techniques known in the art, and those skilled in the art can implement this large language model and multi-role virtual expert service based on the description in this specification.

[0069] Example 4: This embodiment applies to the specific process of generating personalized digital broker images for different users. Users include three typical roles: university researchers who need technology transfer services, founders of startups who need comprehensive entrepreneurial services, and investment institution personnel who need project screening services.

[0070] The data stream generated by the personalized digital broker is as follows: The user management module transmits user registration information, such as industry, identity type, and contact information, to the user profile analysis module.

[0071] The user profile analysis module also receives historical user behavior data from the data management module, such as usage behavior, access records, and operating habits. After integrating and analyzing the registration information with the historical behavior data, it generates a user profile and transmits the profile data to the personalized digital broker generation module.

[0072] The personalized digital agent generation module includes a profile feature encoding submodule, a generative AI image generation submodule, and an image configuration output submodule. The profile feature encoding submodule converts user profiles into feature vectors, the generative AI image generation submodule generates agent images based on the feature vectors, and the image configuration output submodule outputs the generated image configuration to a multimodal interactive display platform for display.

[0073] The multimodal interactive display platform presents the generated agent avatar to users. User feedback, such as satisfaction ratings of the avatar and preferences expressed during interactions, is transmitted back to the user profile analysis module through the data management module for dynamic updates to the profile.

[0074] The process of generating a personalized digital broker is as follows: First, the user profiling analysis module collects multi-dimensional user data: (a) Registration information: including industry sector (such as AI healthcare, intelligent manufacturing, new energy, etc.) and identity type (university researchers, startup founders, investment institution personnel, etc.). (b) Usage behavior: including frequency of function use, service call records, and page access paths; (c) Access records: including login frequency, access time period, and duration of each visit; (d) Operating habits: including preferred interaction methods (voice, text, 3D operation), frequently used function entry points, etc.

[0075] Then, the user profile analysis module uses big data analytics to integrate and analyze the aforementioned multi-dimensional data, extracting users' interests (technical fields they follow, types of information they are interested in), usage habits (preferred interaction methods, commonly used service types), and demand preferences (urgently needed service categories, desired consultation directions), generating structured user profiles, and generating multi-dimensional preference tags for users, including industry tags (such as AI healthcare), startup stage tags (such as seed stage, early stage, growth stage), service demand tags (such as legal needs, financial and tax needs, intellectual property needs), and interaction preference tags (such as voice-first, 3D-first).

[0076] Next, after receiving user profile data, the personalized digital broker generation module converts the multi-dimensional labels in the user profile into high-dimensional feature vectors.

[0077] Then, the generative AI image generation submodule inputs the feature vector into the generative AI model (based on StableDiffusion or a similar architecture, fine-tuned and trained with data from science and technology incubation scenarios) to automatically generate a customized digital agent image that matches the user profile.

[0078] The generated image contains configurations in two dimensions: (a) Appearance characteristics configuration: including image style (professional, approachable, capable, etc.), interactive tone (formal, relaxed, etc.), visual presentation (color tone, clothing style, etc.); (b) Professional capability configuration: including focus areas (configure corresponding professional knowledge weights according to user industry tags), recommended roles (preset virtual expert roles to be activated first according to user demand tags), and service priorities (configure service response strategies according to user usage habits).

[0079] Finally, the image configuration output submodule outputs the generated image configuration to the multimodal interactive display platform for display. At the same time, the system updates interest tags in real time through the data management module based on the user's subsequent behavior changes and feedback data, dynamically adjusts the agent's image configuration, and maintains continuous personalization for each user's image.

[0080] User profiles are generated using a tag-based profile construction method based on multi-source data fusion, which includes the following steps: (a) Data preprocessing: Cleaning (removing invalid data and correcting erroneous data), normalization (converting data of different dimensions to a uniform scale), and feature extraction (extracting meaningful features from the raw data) of the collected raw data. (b) User interest modeling: The TF-IDF-based text feature extraction method is used to extract interest keywords from the user's browsing content, and the collaborative filtering algorithm is combined to discover the user's potential interests; (c) Profile Tag Generation: The extracted features are input into a pre-trained classifier (such as a multi-label classification model based on XGBoost or LightGBM), and the output is a set of user preference tags. The generative AI image generation submodule in the personalized digital broker generation module adopts a Conditional Generative Adversarial Network (ConditionalGAN) or Diffusion Model architecture, which uses the user profile feature vector as a conditional input to generate broker image images that conform to the conditional distribution.

[0081] The model's training data consists of a dataset of digital human images labeled with multi-dimensional tags (style, occupational attributes, perceived age, professional field, etc.). Appearance feature configuration is achieved by controlling the style encoding (StyleCode) and attribute encoding (AttributeCode) of the generated model, while professional competence configuration is achieved by configuring the call weights and recommendation strategies of the knowledge base model. The specific implementation methods described above are all conventional techniques known in the art, and those skilled in the art can implement this personalized digital agent generation method based on the description in this specification.

[0082] Example 5: This embodiment applies to a scenario where a user in the field of materials science at a university uses a digital broker system for technology transfer consultation. The user wants to communicate with the digital broker through various interactive methods, including viewing three-dimensional material structure models, browsing technology transfer data charts, asking questions via voice, and viewing relevant image materials.

[0083] The multimodal interactive display platform consists of five layers, which communicate with each other through a unified data bus and event bus. The 3D low-level rendering engine is located at the bottom layer of the platform, implemented based on WebGL technology, and adopts a B / S architecture. The 3D low-level rendering engine connects the 3D model interaction display layer and the data interaction display layer, providing 3D graphics rendering capabilities to the upper layers.

[0084] The 3D model interactive display layer connects to the underlying 3D rendering engine and is responsible for the loading, rendering, and interactive operation management of the 3D model.

[0085] The data interaction and presentation layer connects the underlying 3D rendering engine and the backend API, and is responsible for the visualization and real-time updates of the data.

[0086] The voice interaction presentation layer connects to the backend speech recognition and speech synthesis services via WebSocket, operates independently of the 3D rendering layer, and communicates with other layers through the event bus.

[0087] The image interaction presentation layer connects to the backend image storage service and CDN service via HTTP and HTTPS protocols, operates independently of other layers, and communicates with other layers through an event bus.

[0088] All five presentation layers are connected to a unified event bus. When a user initiates an interaction in one layer, the event bus distributes the interaction event to other layers that need to respond, thus achieving multimodal collaborative display.

[0089] Taking user interaction with the system as an example, the working process of the multimodal interaction platform is as follows: First, after a user logs into the system, the personalized digital agent generation module generates an academic-style AI digital agent based on the user's profile, and then renders and displays it in the browser using a 3D underlying rendering engine.

[0090] Then, the user says through the voice input function of the voice interaction display layer, "I would like to know about the commercialization prospects of my new material patent." The voice interaction display layer sends the voice data to the voice recognition engine for recognition, and the recognition result is transmitted to the multi-role virtual expert service module for processing via the event bus.

[0091] Next, the multi-role virtual expert service module identified that the request involved a commercialization assessment and activated the virtual entrepreneurship mentor service. The virtual entrepreneurship mentor service invoked the entrepreneurship incubation knowledge base model, combined with the user's profile and new material patent data, to generate a commercialization prospect analysis report, including market analysis data, competitive analysis charts, and commercialization path suggestions.

[0092] Then, the various presentation layers of the multimodal interactive display platform collaboratively output service results: 3D Model Interactive Display Layer: Loads the user's new material molecular structure 3D model (GLTF format), allowing users to rotate the model by dragging and dropping with the mouse, zoom in to view details, and click on atoms to view attribute information.

[0093] Data Interaction and Presentation Layer: Integrates the ECharts chart library, displaying material performance comparison data with line charts, market capacity forecasts with bar charts, and competitive landscape distribution with pie charts. Real-time data push is achieved via WebSocket, and charts dynamically change as data updates. Users can filter charts (e.g., by time range), sort (e.g., by performance indicators), and export data (to Excel or as an image).

[0094] Voice interaction presentation layer: The key conclusions of the analysis report are converted into voice output through a voice synthesis engine, while providing waveform visualization feedback and supporting users to ask follow-up questions through voice.

[0095] Image interactive display layer: Displays relevant achievement transformation case images, patent certificate scans, marketing posters, etc. (JPEG, PNG, SVG formats), supports users to browse in gallery mode, zoom in to view details, highlight key content, and share images via social platforms or email.

[0096] Finally, throughout the entire interaction process, users can continuously interact with the AI ​​digital agent through any of the following methods: voice, mouse clicks on 3D models, clicks on chart data points, browsing images, or other means. Each presentation layer maintains state synchronization through an event bus.

[0097] The 3D underlying rendering engine uses Three.js or Babylon.js as the WebGL rendering library, supporting the loading of 3D model formats such as OBJ, FBX, GLTF, and STL. PBR (Physically Based Rendering) materials are implemented to achieve realistic rendering by setting physical parameters such as material roughness, metallicity, and ambient occlusion. The multi-light source system includes ambient light (providing basic illumination), directional light (simulating sunlight), and point light sources (simulating local lighting). Shadows are implemented using shadow mapping technology.

[0098] The sub-second loading of thousands of buildings is achieved through InstancedRendering and Level of Detail (LOD) technologies. This means distant buildings use low-precision models, while nearby buildings use high-precision models. The data interaction display layer integrates ECharts 5.0 or D3.jsv7 as the chart rendering engine. WebSocket uses Socket.IO or the native WebSocket API for bidirectional communication, and SSE (Server-SentEvents) is used for one-way real-time push from the server to the client. The voice interaction display layer uses the Google Cloud Speech-to-Text API or Microsoft Azure Speech SDK for speech recognition, and the speech synthesis engine uses the Google Cloud Text-to-Speech API or Amazon Polly. Multi-language support is achieved by configuring the engine's language parameters (e.g., languageCode='zh-CN', languageCode='en-US'). The image interaction display layer uses a lazy loading strategy for image loading, meaning images are only loaded when they enter the viewport. CDN uses Alibaba Cloud OSS or AWS CloudFront for accelerated image distribution. The AR functionality is implemented by integrating a WebAR framework (such as 8thWall or AR.js), supporting the overlay of 3D models or images onto the camera feed of the real-world environment. The specific technical choices and implementation methods described above are all conventional techniques known in the art, and those skilled in the art can implement this multimodal interactive display platform based on the description in this specification.

[0099] Example 6: This example uses a complete user interaction lifecycle to demonstrate the entire process from initial user registration to continuous use and system optimization.

[0100] This embodiment involves the collaborative work of all modules of the system, and the connection relationship between the modules is the same as in Embodiment 1.

[0101] The complete interaction method includes the following steps: Step 1: Data Acquisition and User Profile Generation The user registered by scanning a QR code with WeChat in the user management module, and filled in information such as industry (smart manufacturing), identity (co-founder of a startup), and company size (20 people). The user management module then transmitted this registration information to the user profile analysis module.

[0102] In subsequent use, the user profile analysis module continuously collects the user's usage behavior data (such as repeatedly calling the virtual lawyer service to view contract templates and browsing intellectual property-related information), access records (such as logging in 3 times a week, staying for an average of 25 minutes each time, and mainly using it in the evening) and operating habits (preferring voice input to text input).

[0103] The user profile analysis module uses big data analytics to integrate and analyze the above data to generate a user profile, which includes the following tags: industry sector tag "intelligent manufacturing", startup stage tag "early stage", service needs tag "primarily legal needs, secondarily intellectual property needs", and interaction preference tag "voice preferred".

[0104] Step Two: Personalized Digital Broker Generation The personalized digital agent generation module uses a generative AI model to automatically generate a sophisticated and professional customized digital agent avatar based on the user's profile, named "AI Digital Agent." This avatar presents a business-like appearance, uses a formal yet approachable tone in its interaction, and prioritizes virtual lawyer and virtual intellectual property advisor services in its professional capabilities.

[0105] Step 3: Interaction request reception, intent recognition, and multi-role matching The user said through the voice interaction layer: "AI digital broker, please review this cooperation agreement for me. Also, I would like to understand the patent layout strategy in the field of intelligent manufacturing."

[0106] The multi-role virtual expert service module receives the interaction request and calls the generative artificial intelligence large language model to perform semantic analysis and intent recognition.

[0107] The intent recognition results include two service requirements: (a) Contract review falls under the legal field; (b) Patent layout strategy falls under the field of intellectual property. Based on the identification results, the multi-role virtual expert service module simultaneously activates virtual lawyer services and virtual intellectual property advisor services.

[0108] The virtual lawyer service uses the enterprise legal knowledge base model to automatically review the cooperation agreement uploaded by the user, identify risky clauses in the contract (such as unclear intellectual property ownership clauses, overly broad confidentiality obligations, etc.), and generate modification suggestions and risk warnings.

[0109] The virtual intellectual property advisory service draws on a subset of intellectual property from the enterprise's legal knowledge base model and, combined with the characteristics of the intelligent manufacturing field, generates patent layout strategy suggestions, including the application sequence of core patents, the layout direction of peripheral patents, and an analysis of competitors' patent positions.

[0110] Step 4: Multimodal Interaction Output and Service Recommendation The multimodal interactive display platform outputs service results in multiple ways: The 3D model interactive display layer showcases the 3D images of a virtual lawyer and a virtual intellectual property consultant, presented side by side. Users can switch to the corresponding detailed service interface by clicking on either image.

[0111] The data interaction and presentation layer uses a heatmap to show the distribution of contract risk clauses and a bubble chart to show the field distribution of patent layout.

[0112] The voice interaction presentation layer outputs core risk warnings and key suggestions through voice synthesis.

[0113] The interactive image display layer shows images of contract modification marks and schematic diagrams of patent layouts.

[0114] Meanwhile, the intelligent recommendation module recommends relevant services based on the user's profile and the content of this interaction, such as a database of intellectual property infringement cases in the field of intelligent manufacturing, a comprehensive collection of contract templates for intelligent manufacturing enterprises, and other information and services.

[0115] Step 5: Feedback Recording and Iterative Optimization The data management module records all behavioral data of the user during this interaction, including: which contract clause modification suggestions were viewed, which patent layout suggestions were read in depth, the complete record of voice interaction, and the click status of recommended content.

[0116] The data management module integrates this feedback data with the user's existing user profile and updates the user profile. For example, if it finds that the user shows a high interest in patent layout-related content, the system automatically increases the weight of the patent layout tag; if it finds that the user pays close attention to intellectual property ownership clauses after contract review, the system adds a refined intellectual property ownership tag.

[0117] The dataset expansion module performs manual data iteration processing on the updated user data according to the strategies of data construction, data collection, prompt word labeling, data classification, and multi-source data feeding. The processed data is then fed into the generative artificial intelligence large language model for incremental training or fine-tuning to achieve continuous optimization of the model.

[0118] At this point, the system has completed a full closed-loop process, from data collection, profile generation, image customization, intent recognition and multi-role matching, multimodal interaction output to feedback iteration. As the user continues to use the system, the system will continuously optimize the accuracy and personalization of its services.

[0119] The big data analytics technology in step one specifically refers to the analysis method that combines batch processing and stream processing based on a distributed computing framework (such as Apache Spark): batch processing is used for offline analysis of historically accumulated user registration information and long-term behavioral patterns (executed once a day at midnight), stream processing is used for online analysis of users' real-time operational behavior, and user profiles are stored using a three-tier storage architecture of Redis caching for real-time queries + Elasticsearch (for complex retrieval) + HBase (for historical data storage).

[0120] The generative AI model in step two specifically refers to the Conditional Diffusion Model, which has been fine-tuned with data from the science and technology innovation incubation field. The input of the model is a text-based description of the user profile (such as the founder of a startup in the field of intelligent manufacturing, who needs legal and intellectual property services), and the output is a 512×512 resolution image of the person.

[0121] Step 3, "Semantic Analysis and Intent Recognition," employs a few-shot prompting method based on a large language model. By providing 5-10 examples in the prompt template (such as the user saying "help me review the contract," "legal matters," or "activate the virtual lawyer"), the large language model is guided to classify the user input.

[0122] In step four, the multimodal collaborative output achieves state synchronization of each presentation layer through a unified event bus mechanism. When a user performs an operation in one layer (such as clicking on a 3D model), the event bus distributes the event to other layers, triggering corresponding linkage responses (such as the data layer synchronously updating relevant charts and the voice layer synchronously broadcasting relevant information).

[0123] Step five's incremental training or fine-tuning employs the LoRA (Low-Rank Adaptation) parameter-efficient fine-tuning method, updating only a subset of model parameters. This significantly reduces computational resource consumption while ensuring improved model performance. The dataset expansion module's data construction refers to designing the data collection schema and annotation specifications based on business needs; data collection refers to obtaining raw data from system logs, user feedback, and third-party data sources; and prompt word tagging refers to domain experts annotating the data (e.g., labeling user intent categories, service domain categories, etc.). Data classification refers to categorizing the labeled data according to the training requirements of the five sub-models (e.g., legal data is categorized into the training set of the corporate legal knowledge base model); multi-source data feeding refers to mixing data from different data sources according to a predetermined ratio and then inputting it into the model for training. The specific implementation methods described above are all conventional techniques known in the art, and those skilled in the art can implement this complete interactive method based on the description in this specification.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital brokerage system for science and technology innovation incubation based on generative artificial intelligence, characterized in that: include: The user management module supports user registration, login, and user role management. The user profile analysis module, connected to the user management module, is used to collect user profile data through user registration information, usage behavior, access records and operation habits, use big data analysis technology to extract user interests, usage habits and needs preferences to generate user profiles, and dynamically update user preferences and interest tags according to changes in user behavior. The personalized digital agent generation module is connected to the user profile analysis module. It is used to automatically generate a customized digital agent image based on the user profile through a generative AI model, so as to achieve one image per user. The underlying framework for AI services is used to provide underlying support for model training, model management, and model application. Generative AI large language model, deployed on the underlying framework of the AI ​​service, includes at least a startup incubation knowledge base model, a corporate legal knowledge base model, a corporate financial knowledge base model, a corporate tax knowledge base model, and a planning text model; The multi-role virtual expert service module connects the generative artificial intelligence big language model and the personalized digital broker generation module. Based on the user profile and the big language model, it provides users with virtual entrepreneurial mentor services, virtual lawyer services, virtual financial and tax expert services, virtual intellectual property consultant services, and virtual planning text writing expert services. A multimodal interactive display platform, connecting the personalized digital agent generation module and the multi-role virtual expert service module, is used to realize multimodal interaction between users and digital agents through at least two of the following methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction. The intelligent recommendation module connects the user profile analysis module and the multi-role virtual expert service module. It is used to intelligently recommend content and services that users are interested in based on the user profile, and dynamically optimize the recommendation algorithm according to the user's real-time behavior and feedback. The data management module connects the user profile analysis module and the underlying AI service framework. It is used to collect data through multiple channels, clean, store and analyze the collected data to uncover user needs and behavioral patterns.

2. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, The underlying framework of the AI ​​service includes: a data acquisition module for obtaining massive amounts of data from multiple data sources; A network configuration module used to provide high-bandwidth, low-latency network connectivity and load balancing; A service deployment module for enabling rapid deployment and elastic scaling of AI models through containerization technology; A service monitoring module for real-time monitoring and logging of all stages of model training and application; A security and compliance module used to protect data privacy through data encryption and access control; A performance optimization module for improving model inference speed through hardware accelerators.

3. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, In the aforementioned generative artificial intelligence large language model: The startup incubation knowledge base model covers startup theory, business models, and market analysis knowledge; The corporate legal knowledge base model covers legal regulations, compliance requirements, and contract management knowledge. The enterprise financial knowledge base model covers knowledge of financial management, budgeting, and capital operation. The enterprise tax knowledge base model provides information on tax planning, tax declaration, and tax compliance. The proposed text model is used for the generation and optimization of business plans, project reports, and marketing copy.

4. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, In the multi-role virtual expert service module: The virtual entrepreneurship mentor service is used to provide assistance with entrepreneurship direction selection, business model design, market research, competitive analysis, success case analysis, and resource matching. The virtual lawyer service is used to provide legal advice and legal risk assessment on company establishment, contract review, labor law, and intellectual property law. The virtual financial and tax expert service is used to provide financial planning, budget management, fundraising, tax planning, and financial statement analysis. The virtual intellectual property advisory service is used to provide guidance on patent applications, consultation on trademark registration, and assessment of intellectual property infringement risks. The virtual project writing expert service provides guidance on writing business plans, project proposals, and marketing plans.

5. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, The multimodal interactive display platform includes: A 3D underlying rendering engine based on WebGL technology and using a B / S architecture; An interactive data presentation layer used for visualizing data in chart form; A 3D model interactive display layer used to support 3D model loading and rendering operations; A voice interaction presentation layer used to support voice input recognition and voice feedback; And an image interaction display layer to support image display and interactive operations.

6. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system and interaction method according to claim 5, characterized in that, The 3D underlying rendering engine supports loading 3D models in OBJ, FBX, GLTF and STL formats, implements PBR physically based rendering materials, and supports multi-light source lighting effects as well as model rotation, scaling, translation, selection and highlighting. The data interaction presentation layer integrates ECharts or D3.js chart libraries, and uses WebSocket or Server-SentEvents to achieve real-time data push and dynamic chart updates, and provides data filtering, sorting, filtering and data export functions.

7. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 5, characterized in that, The voice interaction presentation layer integrates a speech recognition engine and a speech synthesis engine, and uses natural language processing technology to realize multi-turn dialogue, contextual understanding, and automatic detection and switching of multilingual input and output; The image interactive display layer supports loading images in JPEG, PNG and SVG formats, provides gallery mode, slideshow mode and single image mode display modes, and implements image zooming, dragging, cropping, rotating, annotation and sharing functions.

8. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, It also includes a generative artificial intelligence image model, which is a diffusion model, connected to the multi-role virtual expert service module, and used for chart generation, effect image generation, and PPT generation to assist the output results of the multi-role virtual expert service module.

9. The generative artificial intelligence-based science and technology innovation incubation digital brokerage system according to claim 1, characterized in that, It also includes a dataset expansion module, which is connected to the underlying framework of the AI ​​service. It is used to expand and optimize the dataset by performing manual data iteration through data construction, data collection, prompt word labeling, data classification, and multi-source data feeding strategies.

10. A method for interactive digital brokerage for science and technology innovation incubation based on generative artificial intelligence, employing the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: Step 1: Obtain user data through the user management module, and generate and dynamically update user profiles through the user profile analysis module; Step 2: Based on the user profile, the personalized digital agent generation module calls the generative AI model to automatically generate a customized digital agent image. Step 3: Receive user interaction requests through the multi-role virtual expert service module, identify user intent based on the generative artificial intelligence large language model, and call the corresponding professional model from the startup incubation knowledge base model, corporate legal knowledge base model, corporate financial knowledge base model, corporate tax knowledge base model and planning text model to match and activate the corresponding virtual expert role. Step 4: Through the multimodal interactive display platform, the service results are output to the user using at least two of the following interaction methods: 3D model interaction, data visualization interaction, voice interaction, and image interaction. Service content is also recommended to the user through the intelligent recommendation module. Step 5: Record user feedback and behavioral data through the data management module, update the user profile, and feed the updated data into the generative artificial intelligence large language model for iterative optimization.

Images