Low-threshold AI development system based on automation technology

Abstract

The invention relates to the technical field of artificial intelligence development tools, in particular to a low-threshold AI development system based on an automation technology. According to the technical scheme, the low-threshold AI development system based on the automation technology comprises a data processing and security module, a low-code model building module, a model interpretation and credibility module, a multi-environment deployment module, an intelligent workflow arrangement module and a plug-in ecological and expansion module. Automatic preprocessing is achieved through a multi-modal data analyzer and a dynamic cleaning rule base, data samples are enhanced in combination with a generative adversarial network, an end-to-end automatic assembly line is constructed, manual intervention is reduced, a development team can concentrate on core algorithm optimization, version control is achieved through Git-LFS integration, and the development team can concentrate on data consanguinity tracking. The data change traceability is ensured, and the data processing efficiency is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence development tools, and in particular to a low-threshold AI development system based on automation technology. Background Art

[0002] The core goal of the AI ​​development system is to significantly lower the entry threshold for AI development by introducing a variety of advanced technologies such as automation, modularization, and visualization, so that developers without a deep technical background can easily get started. At the same time, the system is committed to improving the overall efficiency of AI development, shortening project cycles, and reducing development costs, thereby promoting the widespread application and rapid development of artificial intelligence technology. Through this system, users can more conveniently complete the construction, training, deployment, and maintenance of AI models, and achieve efficient management of the entire process of AI projects.

[0003] Existing AI development systems rely on manually configured rules for feature engineering, data cleaning, and other aspects of unstructured data processing. They lack the ability to automate the entire process from data collection to modeling, resulting in long and error-prone development cycles.

[0004] For existing AI development systems, in the processing of unstructured data, feature engineering, data cleaning and other links rely on manually configured rules, and lack the ability to automate the entire process from collection to modeling, resulting in long development cycles and error-prone problems. This solution uses a multimodal data parser and a dynamic cleaning rule library to achieve automated preprocessing, combined with generative adversarial networks to enhance data samples and build an end-to-end automated pipeline. The adaptive crawler engine can automatically identify web page structures and collect data. The dynamic cleaning rule library supports regular expressions, rule engines and custom script extensions to reduce manual intervention. The development team can focus on core algorithm optimization, implement version control through Git-LFS integration, and combine data lineage tracking to ensure that data changes are traceable, effectively improving data processing efficiency. Summary of the Invention

[0005] In order to overcome the existing AI development system, in unstructured data processing, feature engineering, data cleaning and other links rely on manual configuration rules, lacking the ability to automate the entire process from collection to modeling, resulting in long development cycles and prone to errors.

[0006] The technical solution of the present invention is: a low-threshold AI development system based on automation technology, comprising the following modules: Data processing and security module: used for full life cycle management of data and integrated security protection mechanism; Low-code model building module: used to automate model building through a visual interface; Model explanation and trust module: used to improve model decision transparency and compliance; Multi-environment deployment module: used to support seamless cross-platform deployment and resource optimization; Intelligent workflow orchestration module: used to achieve end-to-end business process automation; Plug-in ecosystem and expansion modules: used to build an open and extensible development ecosystem.

[0007] Preferably, the data processing and security module includes: A11: Intelligent data preprocessing unit, including an adaptive crawler engine, a multimodal data parser, and a dynamic cleaning rule library. It is used to automatically collect and clean web page, API, and database data, and dynamically adapt cleaning rules. A12: Privacy protection enhancement unit, including a differential privacy injector, a homomorphic encryption calculation layer, and a sensitive field dynamic desensitizer. It is used to implement dynamic desensitization of sensitive fields and full-link encrypted calculation through differential privacy injection and homomorphic encryption technology; A13: Data version control unit, including the Git-LFS integration interface, data lineage tracker, and incremental update synchronizer. It is used to integrate Git-LFS to implement data lineage tracking and incremental synchronization, and supports version rollback and collaborative auditing.

[0008] Preferably, the data processing and security module comprises the following steps when operating: S11: Automatically identify web page structure, API interface parameters, and database table structure through an adaptive crawler engine, and dynamically adapt data collection rules; S12: After the user configures the data source type, the system automatically generates a collection task, which supports timed triggering or real-time streaming collection; S13: Call the multimodal data parser to convert the format of text, image, and audio unstructured data, and filter out noisy data by combining the dynamic cleaning rule library; S14: Automatically selects parsing templates based on data types. Cleaning rules support regular expressions, rule engines, and custom script extensions. S15: Laplace noise is added during the data collection phase through a differential privacy injector, and combined with the homomorphic encryption computing layer to implement data calculation in the encrypted state; S16: Privacy budget parameters are configured by the user, and encryption keys are dynamically generated and managed by the hardware security module; S17: Dynamic desensitizer for sensitive fields uses regular matching to identify ID card number or mobile phone number fields and replaces the original data using a hash algorithm or fixed masking rules. S18: Desensitization rules support hierarchical control based on role permissions, and audit logs record the entire desensitization operation process; S19: Integrate the Git-LFS interface to implement large file version management, and combine it with the data lineage tracker to record the data processing chain; S110: Automatically submit version snapshots for each data change, and the lineage map supports tracing data sources and processing processes through metadata IDs; S111: The incremental update synchronizer identifies changed data by comparing timestamps or hash values ​​and uses a Merkle tree structure to achieve efficient synchronization. S112: Conflict detection supports manual merging or automatic strategies, and synchronization tasks support breakpoint resumption; S113: GDPR Clause Mapper associates data processing operations with regulatory clauses and generates structured audit logs; S114: The system automatically detects compliance risk points for cross-border data transmission and user authorization, triggers early warnings, and generates rectification suggestions.

[0009] As a preference, the low-code model building module includes: A21: Algorithm component library unit, including pre-trained model market, custom operator dragger and neural architecture search engine, used to provide pre-trained model market and visual operator dragging function, integrated neural architecture search engine; A22: Hyperparameter optimization unit, including the Bayesian optimizer, distributed training scheduler, and early stopping policy controller. It is used to automatically adjust parameters and support early stopping policy control based on Bayesian optimization and distributed training strategies. A23: Model validation unit, including a cross-validation segmenter, an adversarial sample generator, and a robustness evaluation indicator library. It is used to generate adversarial samples and calculate cross-validation indicators to evaluate model robustness and generalization ability.

[0010] Preferably, the low-code model building module includes the following steps when working: S21: Load public models through the pre-trained model market interface and support importing local model libraries; S22: After the user selects a model, the system automatically downloads the weight file, configures the default hyperparameters, and generates an initialization log. S23: The custom operator dragger provides a component library for data processing, feature engineering, and algorithm modules, and supports BPMN standard process connections. S24: The user generates a training pipeline by dragging and dropping components and configuring parameters. The system automatically generates the corresponding Python / JSON code framework. S25: Integrated neural architecture search engine, using reinforcement learning algorithm to automatically generate candidate network structures; S26: The user sets the performance index, and the system outputs the top-K candidate models and compares them visually; S27: Bayesian optimizer combined with a distributed training scheduler to perform hyperparameter search in parallel on a multi-node cluster; S28: After the user defines the search space, the system automatically allocates computing resources, and the early stopping strategy controller monitors the validation set loss and terminates inefficient tasks early. S29: The adversarial sample generator uses algorithms such as FGSM and PGD to construct perturbation data, and combines the robustness evaluation index library to calculate the model's anti-interference ability; S210: After the user configures the attack intensity, the system outputs the adversarial sample set and the accuracy change curve of the model under perturbation; S211: The cross-validation segmenter supports K-fold and stratified K-fold segmentation strategies, and combines the robustness indicator library to evaluate the model generalization error; S212: The system automatically generates a validation set performance report, supporting the sorting of candidate models by accuracy and F1-score indicators; S213: Convert the model to TensorRT or ONNX format using the edge device adapter generator, and generate a Docker image based on the cloud service template library. S214: After the user selects the deployment target, the system automatically compiles the adaptation package and configures the API gateway, generating a deployment status monitoring dashboard.

[0011] Preferably, the model interpretation and trust module includes: A31: Explanatory algorithm unit, including SHAP value calculator, LIME local interpreter and causal reasoning graph generator. It is used to analyze feature importance using SHAP value and LIME algorithm and generate causal reasoning graph to show decision logic. A32: Visualization interaction unit, including a 3D feature space projector, a decision tree dynamic demonstrator, and an attention heat map renderer. It is used to dynamically demonstrate the model decision path by combining 3D feature space projection with attention heat maps. A33: Compliance review unit, including bias detector, GDPR clause mapper and audit log generator, is used to automatically detect bias such as gender / race, map GDPR clauses and generate structured audit logs.

[0012] Preferably, the model interpretation and trust module includes the following steps when working: S31: Call the SHAP value calculator to quantify feature contributions and combine it with the LIME local interpreter to generate a single-sample decision explanation. S32: After the user selects a sample, the system outputs a feature importance bar chart and a neighborhood sample interpretation report generated by LIME; S33: Causal reasoning graph generator builds causal relationships between variables based on structural causal models and displays decision boundaries using a 3D feature space projector. S34: The system automatically generates a causal graph topology and supports adjusting feature dimensions by sliding axes to observe changes in the decision surface. S35: The attention heatmap renderer extracts the attention weights of the model's intermediate layers and combines it with the decision tree dynamic demonstrator to expand the branch nodes layer by layer. S36: After the user clicks on a model layer node, the system highlights the key feature area and plays a decision tree traversal animation; S37: Bias detector counts the correlation between sensitive attributes and prediction results and calculates group fairness indicators; S38: The system outputs bias heatmaps and fairness reports, supports filtering high-bias samples based on thresholds, and triggers manual review. S39: GDPR clause mapper associates the model's personal data processing operations with GDPR Articles 5-21 to generate structured compliance documentation; S310: After the user uploads the data processing agreement, the system automatically marks the terms and conditions and outputs supporting documentation that demonstrates compliance with EU regulatory requirements. S311: The audit log generator records model training, inference, and modification operations, using blockchain technology to store key nodes; S312: Users can search logs by time range or operator, and support exporting audit reports with digital signatures. S313: Collect user feedback on interpretation results through the comment annotation system and update the model in combination with the error pattern learner; S314: After the user marks the sample with inaccurate explanation, the system automatically triggers the model fine-tuning task and generates a comparison report of the improvement effect.

[0013] Preferably, the multi-environment deployment module includes: A41: Deployment configuration unit, including edge device adaptation package, cloud service template library and WebAssembly compilation chain, used to generate edge device, cloud service and WebAssembly adaptation package in one click; A42: Dynamic Optimization Unit, including a quantization compressor, a model pruning policy library, and a hardware-aware scheduler. It is used to achieve optimal resource allocation through quantization compression and model pruning techniques combined with hardware-aware scheduling. A43: Service monitoring unit, including API gateway, QoS indicator collector and automatic scaling controller, is used to collect QoS indicators through API gateway, support automatic scaling and real-time performance warning.

[0014] Preferably, the intelligent workflow orchestration module includes: A51: Visual orchestration unit, including a BPMN 2.0 standard node library, conditional branch triggers, and a cyclic task generator. It is used to drag and drop process nodes based on the BPMN 2.0 standard and integrates conditional branch and cyclic task control. A52: Exception handling unit, including a breakpoint-resume retry mechanism, a manual intervention interface, and an error pattern learner. This unit combines the breakpoint-resume mechanism with error pattern learning to support manual intervention and annotation collaboration. A53: Team collaboration unit, including a real-time collaborative editor, a role permission matrix, and a comment and annotation system. It is used to achieve real-time collaborative editing through Operational Transform technology and configure a refined role permission matrix.

[0015] As a preferred option, the plug-in ecosystem and expansion modules include: A61: Plugin Marketplace unit, including a sandbox runtime environment, dependency conflict detector, and version compatibility checker. It is used to run third-party plugins in isolation in a sandbox environment and automatically detect dependency conflicts and version compatibility. A62: API integration unit, including REST / gRPC interface generator, OAuth2.0 authentication middleware and Webhook notification system, is used to automatically generate REST / gRPC interfaces and integrate OAuth2.0 authentication and Webhook notification systems; A63: Hardware expansion unit, including ROS robot interface, IoT device protocol converter and AR / VR rendering engine. It is used to provide ROS robot interface and IoT protocol converter and support AR / VR rendering engine integration.

[0016] Beneficial effects of the present invention: 1. Existing AI development systems rely on manually configured rules for unstructured data processing, such as feature engineering and data cleaning. They lack full-process automation from acquisition to modeling, resulting in long and error-prone development cycles. This solution uses a multimodal data parser and a dynamic cleaning rule library to automate preprocessing. This solution, combined with generative adversarial networks to enhance data samples, builds an end-to-end automated pipeline. The adaptive crawler engine automatically identifies web page structures and collects data. The dynamic cleaning rule library supports regular expressions, rule engines, and custom script extensions, reducing manual intervention and allowing development teams to focus on core algorithm optimization. Git-LFS integration enables version control, combined with data lineage tracking, ensures traceability of data changes, and effectively improves data processing efficiency. 2. Existing AI development systems lack built-in explanation tools, making it difficult for users to understand model decision logic, hindering their application in key areas and increasing compliance risks. This solution integrates a SHAP value calculator, a LIME local explainer, and a causal inference graph generator to provide a multi-dimensional explanation solution. A three-dimensional feature space projector combined with an attention heat map dynamically displays the model's decision path. A GDPR clause mapper automatically associates regulatory requirements and generates structured compliance reports, significantly enhancing user trust. 3. Existing AI development systems rely on specific cloud service providers. Edge devices, private servers, and other scenarios require secondary development, and resource scheduling lacks dynamic optimization capabilities. This solution designs an edge device adapter package generator and a WebAssembly compilation chain, combined with a hardware-aware scheduler to achieve intelligent resource allocation. Through quantization compression and model pruning technology, the model size is reduced by 80%, while supporting seamless switching between the cloud and the edge, effectively reducing deployment costs and improving resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Shown is a schematic diagram of the framework flow of a low-threshold AI development system based on automation technology of the present invention; Figure 2 Shown is a schematic diagram of the workflow of the data processing and security module of a low-threshold AI development system based on automation technology of the present invention; Figure 3 Shown is a workflow diagram of a low-code model construction module of a low-threshold AI development system based on automation technology of the present invention; Figure 4 What is shown is a schematic diagram of the model interpretation and trusted module workflow of a low-threshold AI development system based on automation technology of the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described below with reference to the accompanying drawings and examples.

[0019] See also Figure 1 The present invention provides an embodiment: a low-threshold AI development system based on automation technology, comprising the following modules: Data processing and security module: used for full life cycle management of data and integrated security protection mechanism; Low-code model building module: used to automate model building through a visual interface; Model explanation and trust module: used to improve model decision transparency and compliance; Multi-environment deployment module: used to support seamless cross-platform deployment and resource optimization; Intelligent workflow orchestration module: used to achieve end-to-end business process automation; Plug-in ecosystem and expansion modules: used to build an open and extensible development ecosystem.

[0020] See also Figure 2-4 In this embodiment, the data processing and security module includes: A11: Intelligent data preprocessing unit, including an adaptive crawler engine, a multimodal data parser, and a dynamic cleaning rule library. It is used to automatically collect and clean web page, API, and database data, and dynamically adapt cleaning rules. A12: Privacy protection enhancement unit, including a differential privacy injector, a homomorphic encryption calculation layer, and a sensitive field dynamic desensitizer. It is used to implement dynamic desensitization of sensitive fields and full-link encrypted calculation through differential privacy injection and homomorphic encryption technology; A13: Data version control unit, including the Git-LFS integration interface, data lineage tracker, and incremental update synchronizer. It is used to integrate Git-LFS to implement data lineage tracking and incremental synchronization, and supports version rollback and collaborative auditing.

[0021] Preferably, the data processing and security module comprises the following steps when operating: S11: Automatically identify web page structure, API interface parameters, and database table structure through an adaptive crawler engine, and dynamically adapt data collection rules; S12: After the user configures the data source type, the system automatically generates a collection task, which supports timed triggering or real-time streaming collection; S13: Call the multimodal data parser to convert the format of text, image, and audio unstructured data, and filter out noisy data by combining the dynamic cleaning rule library; S14: Automatically selects parsing templates based on data types. Cleaning rules support regular expressions, rule engines, and custom script extensions. S15: Laplace noise is added during the data collection phase through a differential privacy injector, and combined with the homomorphic encryption computing layer to implement data calculation in the encrypted state; S16: Privacy budget parameters are configured by the user, and encryption keys are dynamically generated and managed by the hardware security module; S17: Dynamic desensitizer for sensitive fields uses regular matching to identify ID card number or mobile phone number fields and replaces the original data using a hash algorithm or fixed masking rules. S18: Desensitization rules support hierarchical control based on role permissions, and audit logs record the entire desensitization operation process; S19: Integrate the Git-LFS interface to implement large file version management, and combine it with the data lineage tracker to record the data processing chain; S110: Automatically submit version snapshots for each data change, and the lineage map supports tracing data sources and processing processes through metadata IDs; S111: The incremental update synchronizer identifies changed data by comparing timestamps or hash values ​​and uses a Merkle tree structure to achieve efficient synchronization. S112: Conflict detection supports manual merging or automatic strategies, and synchronization tasks support breakpoint resumption; S113: GDPR Clause Mapper associates data processing operations with regulatory clauses and generates structured audit logs; S114: The system automatically detects compliance risk points for cross-border data transmission and user authorization, triggers early warnings, and generates rectification suggestions.

[0022] As a preference, the low-code model building module includes: A21: Algorithm component library unit, including pre-trained model market, custom operator dragger and neural architecture search engine, used to provide pre-trained model market and visual operator dragging function, integrated neural architecture search engine; A22: Hyperparameter optimization unit, including the Bayesian optimizer, distributed training scheduler, and early stopping policy controller. It is used to automatically adjust parameters and support early stopping policy control based on Bayesian optimization and distributed training strategies. A23: Model validation unit, including a cross-validation segmenter, an adversarial sample generator, and a robustness evaluation indicator library. It is used to generate adversarial samples and calculate cross-validation indicators to evaluate model robustness and generalization ability.

[0023] Preferably, the low-code model building module includes the following steps when working: S21: Load public models through the pre-trained model market interface and support importing local model libraries; S22: After the user selects a model, the system automatically downloads the weight file, configures the default hyperparameters, and generates an initialization log. S23: The custom operator dragger provides a component library for data processing, feature engineering, and algorithm modules, and supports BPMN standard process connections. S24: The user generates a training pipeline by dragging and dropping components and configuring parameters. The system automatically generates the corresponding Python / JSON code framework. S25: Integrated neural architecture search engine, using reinforcement learning algorithm to automatically generate candidate network structures; S26: The user sets the performance index, and the system outputs the top-K candidate models and compares them visually; S27: Bayesian optimizer combined with a distributed training scheduler to perform hyperparameter search in parallel on a multi-node cluster; S28: After the user defines the search space, the system automatically allocates computing resources, and the early stopping strategy controller monitors the validation set loss and terminates inefficient tasks early. S29: The adversarial sample generator uses algorithms such as FGSM and PGD to construct perturbation data, and combines the robustness evaluation index library to calculate the model's anti-interference ability; S210: After the user configures the attack intensity, the system outputs the adversarial sample set and the accuracy change curve of the model under perturbation; S211: The cross-validation segmenter supports K-fold and stratified K-fold segmentation strategies, and combines the robustness indicator library to evaluate the model generalization error; S212: The system automatically generates a validation set performance report, supporting the sorting of candidate models by accuracy and F1-score indicators; S213: Convert the model to TensorRT or ONNX format using the edge device adapter generator, and generate a Docker image based on the cloud service template library. S214: After the user selects the deployment target, the system automatically compiles the adaptation package and configures the API gateway, generating a deployment status monitoring dashboard.

[0024] Preferably, the model interpretation and trust module includes: A31: Explanatory algorithm unit, including SHAP value calculator, LIME local interpreter and causal reasoning graph generator. It is used to analyze feature importance using SHAP value and LIME algorithm and generate causal reasoning graph to show decision logic. A32: Visualization interaction unit, including a 3D feature space projector, a decision tree dynamic demonstrator, and an attention heat map renderer. It is used to dynamically demonstrate the model decision path by combining 3D feature space projection with attention heat maps. A33: Compliance review unit, including bias detector, GDPR clause mapper and audit log generator, is used to automatically detect bias such as gender / race, map GDPR clauses and generate structured audit logs.

[0025] Preferably, the model interpretation and trust module includes the following steps when working: S31: Call the SHAP value calculator to quantify feature contributions and combine it with the LIME local interpreter to generate a single-sample decision explanation. S32: After the user selects a sample, the system outputs a feature importance bar chart and a neighborhood sample interpretation report generated by LIME; S33: Causal reasoning graph generator builds causal relationships between variables based on structural causal models and displays decision boundaries using a 3D feature space projector. S34: The system automatically generates a causal graph topology and supports adjusting feature dimensions by sliding axes to observe changes in the decision surface. S35: The attention heatmap renderer extracts the attention weights of the model's intermediate layers and combines it with the decision tree dynamic demonstrator to expand the branch nodes layer by layer. S36: After the user clicks on a model layer node, the system highlights the key feature area and plays a decision tree traversal animation; S37: Bias detector counts the correlation between sensitive attributes and prediction results and calculates group fairness indicators; S38: The system outputs bias heatmaps and fairness reports, supports filtering high-bias samples based on thresholds, and triggers manual review. S39: GDPR clause mapper associates the model's personal data processing operations with GDPR Articles 5-21 to generate structured compliance documentation; S310: After the user uploads the data processing agreement, the system automatically marks the terms and conditions and outputs supporting documentation that demonstrates compliance with EU regulatory requirements. S311: The audit log generator records model training, inference, and modification operations, using blockchain technology to store key nodes; S312: Users can search logs by time range or operator, and support exporting audit reports with digital signatures. S313: Collect user feedback on interpretation results through the comment annotation system and update the model in combination with the error pattern learner; S314: After the user marks the sample with inaccurate explanation, the system automatically triggers the model fine-tuning task and generates a comparison report of the improvement effect.

[0026] Preferably, the multi-environment deployment module includes: A41: Deployment configuration unit, including edge device adaptation package, cloud service template library and WebAssembly compilation chain, used to generate edge device, cloud service and WebAssembly adaptation package in one click; A42: Dynamic Optimization Unit, including a quantization compressor, a model pruning policy library, and a hardware-aware scheduler. It is used to achieve optimal resource allocation through quantization compression and model pruning techniques combined with hardware-aware scheduling. A43: Service monitoring unit, including API gateway, QoS indicator collector and automatic scaling controller, is used to collect QoS indicators through API gateway, support automatic scaling and real-time performance warning.

[0027] Preferably, the intelligent workflow orchestration module includes: A51: Visual orchestration unit, including a BPMN 2.0 standard node library, conditional branch triggers, and a cyclic task generator. It is used to drag and drop process nodes based on the BPMN 2.0 standard and integrates conditional branch and cyclic task control. A52: Exception handling unit, including a breakpoint-resume retry mechanism, a manual intervention interface, and an error pattern learner. This unit combines the breakpoint-resume mechanism with error pattern learning to support manual intervention and annotation collaboration. A53: Team collaboration unit, including a real-time collaborative editor, a role permission matrix, and a comment and annotation system. It is used to achieve real-time collaborative editing through Operational Transform technology and configure a refined role permission matrix.

[0028] As a preferred option, the plug-in ecosystem and expansion modules include: A61: Plugin Marketplace unit, including a sandbox runtime environment, dependency conflict detector, and version compatibility checker. It is used to run third-party plugins in isolation in a sandbox environment and automatically detect dependency conflicts and version compatibility. A62: API integration unit, including REST / gRPC interface generator, OAuth2.0 authentication middleware and Webhook notification system, is used to automatically generate REST / gRPC interfaces and integrate OAuth2.0 authentication and Webhook notification systems; A63: Hardware expansion unit, including ROS robot interface, IoT device protocol converter and AR / VR rendering engine. It is used to provide ROS robot interface and IoT protocol converter and support AR / VR rendering engine integration.

[0029] Example 1: Development of a smart government AI assistant based on the Nova platform Implementation background: A municipal government planned to build an intelligent government consultation system, but faced three major technical bottlenecks: (1) Fragmented data processing: Unstructured data such as historical consultation records and policy documents require manual labeling and cleaning, which is inefficient; (2) High threshold for model development: Business personnel lack algorithm foundation and rely on third-party manufacturers to customize models, which takes up to 3 months; (3) Difficult deployment and adaptation: It needs to support government cloud, edge self-service machines and mobile apps at the same time, and the existing platform is not cross-platform compatible.

[0030] Implementation steps: S41: Automatically crawls government website policy documents and historical consultation records through an adaptive crawler engine, dynamically adapting the web page structure and API interface; S42: After the user configures the data source type, the system generates a scheduled collection task and supports hourly incremental updates; S43: Multimodal data parser converts unstructured data into structured formats: speech-to-text conversion uses the Wav2Vec2 model, and image data uses ResNet-18 to extract features; S44: The dynamic desensitizer for sensitive fields performs hash encryption on ID card numbers and address information, and the desensitization rules are controlled by the "Government Data Security Level"; S45: Manage data versions through the Git-LFS interface, generate a lineage map for each change, and support tracing data sources by "policy document ID"; S46: GDPR clause mapper automatically detects cross-border data transfer scenarios, triggers compliance alerts, and generates corrective action suggestions; S47: Load the Chinese government dialogue model from the pre-trained model market and automatically configure the default hyperparameters; S48: Import the local policy terminology library, build the "Policy Question Answering" process through the custom operator dragger, and connect the "Text Classification" and "Entity Extraction" components; S49: A Bayesian optimizer searches for hyperparameters in parallel on a 4-node GPU cluster, and an early stopping controller terminates inefficient tasks when the validation set loss does not decrease for three consecutive rounds. S410: The adversarial sample generator simulates user input typos to test the robustness of the model, and the accuracy drop is controlled within 5%; S411: Select the "Edge Device" deployment target. The system automatically generates a TensorRT model package and Docker image, and exposes a RESTful interface through the API gateway. S412: When a user queries "One-child Subsidy Policy," the SHAP value calculator shows that the "Household Registration Type" feature has the highest contribution, and the causal reasoning diagram displays the decision path. S413: The bias detector scans the training data and finds that the sample size for "urban-rural differences" is insufficient, triggering manual review and supplementation of data; S414: Generates a browser-side inference package through the WebAssembly compilation chain, which can be directly called by government apps; S415: The hardware-aware scheduler dynamically allocates computing power to edge devices, keeping CPU utilization below 80% and response time <300ms.

[0031] Comparison table: Comparison Dimension Existing technology platform Nova Solution Improvement Development efficiency Manual intervention accounts for 60%, and the cycle is 3 months Automation rate 85%, cycle 10 days Efficiency increased by 80% Model credibility Black box model, user trust <40% Visual explanation covers 95% of decision paths Trust increased to 85% Cross-platform costs Dependence on cloud vendors, annual costs ≥ 500,000 yuan Unified interface + dynamic optimization, annual fee ≤ 200,000 yuan Cost reduction of 60% Compliance risks Manual audit, vulnerability response time > 7 days Blockchain evidence storage + automatic warning, response time < 2 hours 90% reduction in risk Teamwork Version conflicts occur frequently and communication costs are high Real-time collaborative editing + role permission matrix Development cycle shortened by 50% Implementation summary: Using the Nova platform, the city government launched an intelligent government system within three months, covering the entire consultation, processing, and feedback process. The system handles an average of 12,000 requests per day with an accuracy rate of 92%, reducing the workload of human agents by 70%.

[0032] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A low-threshold AI development system based on automation technology; characterized by: It consists of the following modules: Data processing and security module: used for full life cycle management of data and integrated security protection mechanism; Low-code model building module: used to automate model building through a visual interface; Model explanation and trust module: used to improve model decision transparency and compliance; Multi-environment deployment module: used to support seamless cross-platform deployment and resource optimization; Intelligent workflow orchestration module: used to achieve end-to-end business process automation; Plug-in ecosystem and expansion modules: used to build an open and extensible development ecosystem.

2. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: The data processing and security module includes: A11: Intelligent data preprocessing unit, including an adaptive crawler engine, a multimodal data parser, and a dynamic cleaning rule library, is used to automatically collect and clean web page, API, and database data, and dynamically adapt cleaning rules. A12: Privacy protection enhancement unit, including a differential privacy injector, a homomorphic encryption calculation layer, and a sensitive field dynamic desensitizer. It uses differential privacy injection and homomorphic encryption technology to achieve dynamic desensitization of sensitive fields and full-link encrypted calculation. A13: Data version control unit, including Git-LFS integration interface, data lineage tracker and incremental update synchronizer, is used to integrate Git-LFS to implement data lineage tracking and incremental synchronization, and supports version rollback and collaborative auditing.

3. The low-threshold AI development system based on automation technology according to claim 2, characterized in that: The data processing and security module includes the following steps when it is working: S11: Automatically identify web page structure, API interface parameters and database table structure through the adaptive crawler engine, and dynamically adapt data collection rules; S12: After the user configures the data source type, the system automatically generates a collection task, supporting timed triggering or real-time streaming collection; S13: Call the multimodal data parser to convert the format of text, image and audio unstructured data, and filter out noisy data by combining the dynamic cleaning rule library; S14: Automatically select parsing templates based on data types, and cleaning rules support regular expressions, rule engines, and custom script extensions; S15: Laplace noise is added during the data collection phase through a differential privacy injector, and the data calculation in the encrypted state is realized in combination with the homomorphic encryption computing layer; S16: Privacy budget parameters are configured by the user, and encryption keys are dynamically generated and managed by the hardware security module; S17: The dynamic desensitizer for sensitive fields identifies ID card number or mobile phone number fields based on regular matching and replaces the original data with a hash algorithm or fixed masking rules; S18: Desensitization rules support hierarchical control based on role permissions, and audit logs record the entire process of desensitization operations; S19: Integrate the Git-LFS interface to implement large file version management, and combine the data lineage tracker to record the data processing link; S110: Automatically submit version snapshots for each data change, and the lineage map supports tracing the data source and processing process through metadata ID; S111: The incremental update synchronizer identifies changed data by comparing timestamps or hash values ​​and uses a Merkle tree structure to achieve efficient synchronization; S112: Conflict detection supports manual merging or automatic strategies, and synchronization tasks support breakpoint resumption; S113: GDPR Clause Mapper associates data processing operations with regulatory clauses and generates a structured audit log; S114: The system automatically detects compliance risk points for cross-border data transmission and user authorization, triggers early warnings, and generates rectification suggestions.

4. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: Low-code model building modules include: A21: Algorithm component library unit, including pre-trained model market, custom operator dragger and neural architecture search engine, used to provide pre-trained model market and visual operator dragging function, integrated neural architecture search engine; A22: Hyperparameter optimization unit, including a Bayesian optimizer, a distributed training scheduler, and an early stopping strategy controller. It is used to automatically adjust parameters and support early stopping strategy control based on Bayesian optimization and distributed training strategies. A23: Model validation unit, including a cross-validation segmenter, an adversarial sample generator, and a robustness evaluation indicator library, used to generate adversarial samples and calculate cross-validation indicators to evaluate model robustness and generalization ability.

5. The low-threshold AI development system based on automation technology according to claim 4, characterized in that: The low-code model building module works by: S21: Load public models through the pre-trained model market interface and support importing local model libraries; S22: After the user selects the model, the system automatically downloads the weight file and configures the default hyperparameters, and generates an initialization log; S23: The custom operator dragger provides a library of data processing, feature engineering, and algorithm modules, and supports BPMN standard process connections. S24: The user generates a training pipeline by dragging and dropping components and configuring parameters, and the system automatically generates the corresponding Python / JSON code framework; S25: Integrated neural architecture search engine, using reinforcement learning algorithm to automatically generate candidate network structures; S26: The user sets the performance index, and the system outputs the top-K candidate models and compares them visually; S27: Bayesian optimizer combined with distributed training scheduler to perform hyperparameter search in parallel on a multi-node cluster; S28: After the user defines the search space, the system automatically allocates computing resources, and the early stopping strategy controller monitors the validation set loss and terminates inefficient tasks early; S29: The adversarial sample generator uses algorithms such as FGSM and PGD to construct perturbation data, and combines the robustness evaluation index library to calculate the model's anti-interference ability; S210: After the user configures the attack intensity, the system outputs the adversarial sample set and the accuracy change curve of the model under perturbation; S211: The cross-validation segmenter supports K-fold and stratified K-fold segmentation strategies, and combines the robustness indicator library to evaluate the model generalization error; S212: The system automatically generates a validation set performance report, supporting the sorting of candidate models by accuracy and F1-score indicators; S213: Convert the model into TensorRT or ONNX format through the edge device adaptation package generator, and generate a Docker image in combination with the cloud service template library; S214: After the user selects the deployment target, the system automatically compiles the adaptation package and configures the API gateway, generating a deployment status monitoring dashboard.

6. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: The model interpretation and trust module includes: A31: Explanatory algorithm unit, including SHAP value calculator, LIME local interpreter and causal reasoning graph generator, is used to analyze feature importance using SHAP value and LIME algorithm and generate causal reasoning graph to show decision logic; A32: Visualization interaction unit, including a 3D feature space projector, a decision tree dynamic demonstrator, and an attention heat map renderer, used to dynamically demonstrate the model decision path by combining 3D feature space projection with attention heat maps; A33: Compliance review unit, including bias detector, GDPR clause mapper and audit log generator, is used to automatically detect bias such as gender / race, map GDPR clauses and generate structured audit logs.

7. The low-threshold AI development system based on automation technology according to claim 6, characterized in that: The model interpretation and trust module works by: S31: Call the SHAP value calculator to quantify feature contributions and combine it with the LIME local interpreter to generate a single-sample decision explanation; S32: After the user selects a sample, the system outputs a feature importance bar chart and a neighborhood sample interpretation report generated by LIME; S33: The causal reasoning graph generator constructs causal relationships between variables based on the structural causal model and displays the decision boundary in combination with the three-dimensional feature space projector; S34: The system automatically generates a causal graph topology and supports adjusting feature dimensions by sliding the axis to observe changes in the decision surface; S35: The attention heat map renderer extracts the attention weights of the model's middle layer and combines it with the decision tree dynamic demonstrator to expand the branch nodes layer by layer; S36: After the user clicks on the model layer node, the system highlights the key feature area and plays the decision tree traversal animation; S37: The bias detector counts the correlation between sensitive attributes and prediction results and calculates the group fairness index; S38: The system outputs bias heatmaps and fairness reports, supports filtering high-bias samples by threshold, and triggers manual review; S39: GDPR clause mapper associates the model's operations for processing personal data with GDPR Articles 5-21 to generate structured compliance documentation; S310: After the user uploads the data processing agreement, the system automatically marks the terms and conditions and outputs supporting documentation that it complies with EU regulatory requirements; S311: The audit log generator records model training, inference, and modification operations, using blockchain technology to store key nodes; S312: Users can search logs by time range or operator conditions, and support exporting audit reports with digital signatures; S313: Collect user feedback on the interpretation results through the comment annotation system and update the model in combination with the error pattern learner; S314: After the user marks the sample with inaccurate explanation, the system automatically triggers the model fine-tuning task and generates an improvement effect comparison report.

8. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: The multi-environment deployment module includes: A41: Deployment configuration unit, including edge device adaptation package, cloud service template library and WebAssembly compilation chain, used to generate edge device, cloud service and WebAssembly adaptation package in one click; A42: Dynamic Optimization Unit, including a quantization compressor, a model pruning policy library, and a hardware-aware scheduler. It is used to achieve optimal resource allocation through quantization compression and model pruning techniques combined with hardware-aware scheduling. A43: Service monitoring unit, including API gateway, QoS indicator collector and automatic scaling controller, is used to collect QoS indicators through API gateway, support automatic scaling and real-time performance warning.

9. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: The intelligent workflow orchestration module includes: A51: Visual orchestration unit, including a BPMN 2.0 standard node library, conditional branch triggers, and a cyclic task generator. It is used to drag and drop process nodes based on the BPMN 2.0 standard and integrates conditional branch and cyclic task control. A52: Exception handling unit, including a breakpoint-resume retry mechanism, a manual intervention interface, and an error pattern learner. This unit combines the breakpoint-resume mechanism with error pattern learning to support manual intervention and annotation collaboration. A53: Team collaboration unit, including a real-time collaborative editor, a role permission matrix, and a comment and annotation system. It is used to achieve real-time collaborative editing through Operational Transform technology and configure a refined role permission matrix.

10. The low-threshold AI development system based on automation technology according to claim 1, characterized in that: The plug-in ecosystem and expansion modules include: A61: Plug-in market unit, including sandbox runtime environment, dependency conflict detector and version compatibility checker, is used to run third-party plug-ins in isolation through the sandbox environment and automatically detect dependency conflicts and version compatibility; A62: API integration unit, including REST / gRPC interface generator, OAuth2.0 authentication middleware and Webhook notification system, is used to automatically generate REST / gRPC interfaces and integrate OAuth2.0 authentication and Webhook notification systems; A63: Hardware expansion unit, including ROS robot interface, IoT device protocol converter and AR / VR rendering engine, used to provide ROS robot interface and IoT protocol converter, and support AR / VR rendering engine integration.

Images