agents · LangGraph · FastAPI
Underwriter Agent
Multi-agent mortgage underwriting demo — six LangGraph agents collaborate to render APPROVED / CONDITIONAL / DENIED with risk score and decision memo. Live SSE stream on Hugging Face Spaces.
Overview
A multi-agent mortgage underwriting demo. Six LangGraph agents — Credit, Income, Asset, Collateral, Critic, Decision — collaborate over a synthetic applicant profile through a Supervisor router and render APPROVED / CONDITIONAL_APPROVAL / DENIED with a risk score and decision memo. A live SSE stream lights up the Mermaid workflow diagram as each agent thinks.
What it does
- LangGraph workflow with Supervisor routing across six domain agents.
- RAG retrieval over an underwriting policy PDF, baked into a committed Chroma store (no HF Pro tier required).
- FastAPI backend with SSE streaming; vanilla-JS frontend animating the agent graph.
- PII sanitization and a pure-domain core package with zero web-framework imports.
- Bring-your-own-key: OpenAI key used per-request, never stored. ~$0.15 per full run.
Live demo
→ huggingface.co/spaces/alanvaa06/underwriter-agent
Stack
LangGraph · LangChain · FastAPI · Pydantic v2 · Chroma · OpenAI (gpt-4o) · Mermaid · Tailwind CSS.
Status
Open source (MIT). Productionized from a Johns Hopkins / GreatLearning Agentic AI course lab.