AgentAudit

I built AgentAudit because I kept running into the same frustrating issue with my RAG agents: silent failures. You know the scenario. The retrieval works perfectly, the context chunks are correct, but the LLM still confidently hallucinates a specific detail that isn't in the source text. You often don't find out until a user reports it. I tried fixing this with extensive prompt engineering, but I found that prompts alone are not 100% reliable at scale. I needed a deterministic way to verify answers before they reached the user. That is why I built AgentAudit. It acts as a middleware API that sits between your agent and your frontend. You send it the retrieved context and the generated answer, and it runs a strict logic check to ensure every claim is grounded in the provided text. If it detects a hallucination or a citation error, it flags the response immediately. The Tech Stack I know the AI ecosystem is heavy on Python, but I built this entirely in TypeScript and Node.js because I wanted a strictly typed, reliable backend. Under the hood, it uses PostgreSQL with pgvector for semantic comparison. For Developers I designed this to be framework-agnostic. Since it is just a REST API, you can plug it into LangChain, Flowise, or your custom Python backend without rewriting your agent. It is fully open source for those who want to self-host. For those who just want to test the groundedness check quickly, I have deployed a free tier on RapidAPI that handles the infrastructure for you. I would love to hear how you are currently handling trust and safety in your pipelines. Is a "Judge" model overkill, or necessary for production? Let me know what you think.