mcaleb808/underwrite-ai

AI-powered health insurance underwriting for the Rwandan market # UnderwriteAI AI-assisted health insurance underwriting for the Rwandan market. A multi-agent LangGraph pipeline parses medical PDFs, scores risk against an underwriting manual, drafts a verdict with cited rules, and runs an adversarial fairness critic — then hands off to a human underwriter who can approve, modify, or re-evaluate before an email goes out. ## The problem Underwriting health insurance for a typical Rwandan applicant is a document-heavy, judgment-heavy job: read the medical PDFs, weigh comorbidities against age and BMI, look up the right rules in a binder of underwriting guidelines, decide what to load on the premium, and draft a customer email that's empathetic without leaking internal rule IDs. UnderwriteAI does the mechanical 80% — parsing, scoring, rule retrieval, draft + critic, customer email — and surfaces every step to a senior underwriter for the final call. ## Quick start ```bash # 1. clone + env git clone github.com && cd underwrite-ai cp apps/api/.env.example apps/api/.env # fill OPENROUTER_API_KEY + OPENAI_API_KEY cp apps/web/.env.example apps/web/.env.local # 2. install cd apps/api && uv sync && cd ../.. cd apps/web && pnpm install && cd ../.. # 3. seed Chroma with the underwriting manual make seed # 4. start both services make api # terminal 1 → localhost make web # terminal 2 → localhost ``` Open localhost, pick a seed applicant, and watch the pipeline run. ## How it works ```mermaid flowchart LR Browser["Next.js dashboard"] API["FastAPI (routes · orchestrator · SSE)"] Graph["LangGraph pipeline (5 agents)"] DB[("SQLite")] Chroma[("Chroma")] Email["Email provider"] Browser -- "POST /applications · GET /events (SSE)" --> API API -- "ainvoke · astream" --> Graph Graph --> Chroma API --> DB API --> Email ``` Inside the graph: ```mermaid flowchart LR START((start)) --> doc_parser doc_parser --> risk_assessor doc_parser --> guidelines_rag risk …