PolicyChat AI

PolicyChat AI is a prototype that turns dense policy, legal, and compliance documents into something you can just ask. It was built as a proof of concept — to see how far a retrieval-grounded LLM could go on real, messy documents.

The problem

Nobody reads a 60-page policy PDF to answer one question. I wanted a system that could ingest complex documents — PDF, DOCX, PPTX — and answer questions about them accurately, with enough context to be trusted for legal and policy content.

What I built

• Grounded answers — used Google’s Gemini API with retrieval so responses stay tied to the source document instead of hallucinating.
• Real document ingestion — a pipeline that parses and understands mixed formats (PDF/DOCX/PPTX), not just plain text.
• Persistent, personal sessions — Supabase authentication for secure sign-in, so every user keeps their own multi-session chat history.
• Production shape — FastAPI backend, React frontend, PostgreSQL, containerized with Docker and deployable to GCloud/AWS.

As a prototype, the goal was to prove the hard parts work end to end — grounded retrieval, mixed-format ingestion, and authenticated persistence — and to learn what a production version would need next.