// the find
Ricky-7-Yan/intelligent-audit-system
🤖 AutoAudit--智能审计决策系统 Python FastAPI License 基于大语言模型的智能审计平台 | 集成知识图谱、RAG、强化学习等前沿技术 功能特性 • 快速开始 • 技术架构 • 文档 📋 项目简介 智能审计决策系统是一个基于大语言模型（LLM）的智能审计平台，集成了知识图谱、RAG检索增强生成、强化学习等前沿技术，为审计工作提供智能化支持。 🎯 核心价值: 突破传统审计工具局限，支持复杂业务逻辑的深度推理
AutoAudit is a Python/FastAPI platform that wraps an LLM (Alibaba's Qwen via API) with RAG, a Neo4j knowledge graph, and some training tooling to answer audit-related questions against standards like COBIT, ISO27001, and SOX. It's aimed at auditors or compliance teams in Chinese-speaking organizations who want an AI assistant they can self-host. The codebase is a working prototype, not a production system.
The stack is reasonable for the problem: FastAPI for the API layer, Neo4j for the standards knowledge graph, and a vector store for RAG is a sensible combination. The knowledge graph seeded with actual audit standards (COBIT, SOX, ISO27001) is the most concrete value-add — it gives the LLM something to reason over beyond generic web training data. The training pipeline supporting LoRA fine-tuning and RLHF is a real differentiator for teams that want domain-specific behavior rather than a generic chatbot wrapper. Multi-turn session context in the chat agent is handled, which matters for the back-and-forth nature of real audit work.
The 'performance metrics' in the README — 85% accuracy, 99.9% uptime, 100 concurrent users — are made up; there are no benchmarks, no eval datasets, and no load test results anywhere in the repo. The entire repo is one developer wide: one file per module (agents/audit_agent.py, rag/agentic_rag.py, training/training_pipeline.py), no tests, and no migrations for the MySQL or Neo4j schemas beyond init scripts you run once. The knowledge graph is only as good as what was manually loaded into Neo4j at setup time — there's no documented process for keeping it current with standard revisions. The repo is primarily in Chinese, which limits the contributor pool and makes auditing the logic itself harder for non-Chinese teams considering adoption.