A 7-agent pipeline that processes a home-loan application end to end and hands the human underwriter a decision-ready dashboard. No agent ever approves or rejects autonomously — every credit call stays with the manager.
Dashboard preview — capture instructions in
docs/SCREENSHOT.md. Once captured, the image lives atdocs/dashboard.pngand renders here:
Indian retail mortgage underwriting is a paper-heavy, multi-system workflow: an underwriter has to chase KYC matches across Aadhaar and PAN, sanity-check income against bank statements and Form 16, validate property title with the sub-registrar, and reconcile everything against RBI policy on FOIR and LTV. Most of that is patterning, not judgement. This project pushes the patterning to a sequence of focused LLM agents and reserves the judgement for a human at a clean dashboard.
Two design choices worth flagging up front:
A cost gate at step 2b. Before any paid third-party API is hit (Aadhaar / PAN / CIBIL / land registry), a consistency-check agent decides go/no-go on the document set. Applications with fundamental inconsistencies are routed back to the customer for corrections instead of burning API spend.
A human-in-the-loop terminus. Step 6 compiles a structured dashboard — risk level, key positives, key concerns, conditions for approval, data gaps — but the recommendation is explicitly preliminary. The manager owns the decision; the agents own the prep.
| # | Agent | Job |
|---|---|---|
| 1 | DocumentCollector | Classify and catalogue uploaded files |
| 2 | DocumentAuthenticator | Per-document authenticity check |
| 2b | ConsistencyChecker | Cross-document checks · go/no-go cost gate |
| 3 | KYCVerifier | Aadhaar + PAN verification |
| 4 | FinancialVerifier | CIBIL score, FOIR, Account Aggregator |
| 5 | PropertyVerifier | Title, encumbrance, LTV vs RBI cap |
| 6 | DashboardCompiler | Synthesise into underwriter dashboard |
| 7 | CustomerComms | Draft customer follow-ups |
Prerequisites: Python 3.11+, an Anthropic API key.
git clone https://github.com/Dev251288/underwriting-agent.git
cd underwriting-agent
pip install -e ".[dev]"
cp .env.example .env # set ANTHROPIC_API_KEY in .env
python main.py seed # populate work queue with 5 demo apps
python main.py serve # dashboard at http://127.0.0.1:8000
python main.py demo # full pipeline on the built-in sample
pytest # all tests run without an API keyEvery agent calls Claude with tool_choice={"type": "tool", "name": "..."},
forcing a single structured tool call whose input_schema is the agent's
output contract. System prompts are marked cache_control: {"type": "ephemeral"}
so they are served from cache across repeated runs within the TTL.
| Layer | Choice |
|---|---|
| Agents | Anthropic Claude via anthropic Python SDK |
| API server | FastAPI + Uvicorn |
| Dashboard | Single-page HTML + Tailwind (CDN) + vanilla JS |
| Data models | Pydantic v2 |
| Storage | Per-application JSON in outputs/ |
Production integrations live in src/mocks/ as deterministic stand-ins for:
UIDAI Aadhaar, NSDL PAN, TransUnion CIBIL, RBI Account Aggregator, State
Land Registry. Swap each mock for a real HTTP client to go live — the agent
contracts above don't need to change.
KYC follows RBI KYC Master Directions and the PAN–Aadhaar linking mandate. Credit eligibility uses a CIBIL floor of 650 (with 600–649 routed to Credit Committee, below 600 a hard reject) and a FOIR ceiling of 50% per RBI guidance. LTV ceilings follow the RBI circular: 90% on loans up to ₹30L, 80% on ₹30–75L, 75% above ₹75L; breaches within 5pp of the cap escalate to Credit Committee, beyond 5pp hard-reject. Property records are matched against sub-registrar and RERA registrations.
underwriting-agent/
├── main.py # CLI: demo | serve | seed
├── src/
│ ├── orchestrator.py # 7-step pipeline coordinator
│ ├── agents/ # one file per agent
│ ├── mocks/ # mock external APIs
│ ├── models/ # Pydantic data models
│ └── api/ # FastAPI dashboard
├── tests/ # pytest, no API key needed
├── docs/
│ ├── architecture.svg # README hero
│ └── SCREENSHOT.md # how to capture dashboard.png
└── outputs/ # one JSON per processed application
Tracked in summary.md. Next four items: request-more-info workflow with a
fixed customer SLA, work-queue landing page, exception-only Credit Committee
escalation, and a dashboard redesign that collapses duplicate information into
an expandable scorecard.
After pushing, set these in Settings → General → Topics to make the repo discoverable:
agentic-ai claude anthropic multi-agent human-in-the-loop
fintech mortgage underwriting india rbi python fastapi pydantic
MIT. See LICENSE.