projects / 2026 / Built at work
Credit Helper Agent & policy assistant
Multi-agent RAG for credit and policy questions at AYE Finance, with guardrails, private document chat, and desktop and WhatsApp clients.
What it is
An assistant employees can ask about credit and internal policy. Answers are generated from the policy documents rather than from the model's memory. It has two parts that share one backend:
- Credit Helper Agent: a multi-agent RAG system (LangGraph, AWS Bedrock, Qdrant) over credit and policy documents, with layered prompt-injection and output-safety defences: Bedrock Guardrails plus an LLM-based validation step.
- My Docs: a NotebookLM-style assistant for an employee's own uploaded documents, with citation-grounded answers and data-retention controls.
Shape of it
- Backend: FastAPI, PostgreSQL on RDS, Qdrant for vector search, AWS Bedrock for generation.
- Two modes: an organisation-wide policy chat, scoped by department, and a private "My Docs" mode where uploaded files live in a session and are purged after a time limit.
- Access: employees are matched against an employee table and get a JWT; the admin portal is role-gated through directory login.
- Clients: a web admin portal, an Electron tray app with local file search, and a WhatsApp bridge (Gupshup).
The design notes are in Notes on a policy RAG assistant.