Claude API Lab — Commercial Loan Underwriting Agent

Reference implementation for the Claude API Lab workshop (Module 2.2). An agentic demo that underwrites a small-business loan application end-to-end: Claude drives a tool-use loop that parses a borrower package, validates it against credit policy, computes credit ratios, and produces a structured creditworthiness assessment — no hardcoded step sequence, the model decides which tool to call next.

Branch convention: each industry has an <industry>-solution reference branch and an <industry> fill-in-the-blank starter (e.g. finance-solution + finance, healthcare-solution + healthcare). This main branch currently mirrors finance-solution (the loan-underwriting reference) and is slated to become a distinct domain long-term (see the project plan). Workshop content lives in the cc-workshop platform.

Built by Provectus.

Architecture

React 19 Frontend  ◄──SSE──►  FastAPI Backend  ◄──tool use──►  Claude Sonnet 4.6

Layer	Stack
Frontend	React 19, TypeScript, Vite, MUI, Zustand (SSE streaming UI)
Backend	FastAPI, Pydantic, SSE (sse-starlette), pandas
AI	Claude Sonnet 4.6 via the Anthropic Messages API with tool use

The agent loop lives in backend/app/services/agent_runner.py: call Claude → dispatch each tool_use block → append tool_results → loop until stop_reason == "end_turn" (capped at 25 iterations).

The four tools

Registered in backend/app/tools/registry.py (explicit list — add your module there):

1. parse_loan_package — folder (profile.json + financials.csv) → canonical LoanApplication.
2. validate_loan_application — applies schemas/validation_rules.json → {valid, errors, warnings}.
3. compute_credit_ratios — DSCR, current ratio, debt-to-equity, gross-margin trend, 3-yr revenue CAGR, LTV.
4. assess_creditworthiness — the final verdict (recommendation, pricing, risks, confidence, reasoning). Its output becomes the pipeline assessment.

System prompt: backend/app/prompts/credit_analyst.md. Canonical schema: backend/app/schemas/canonical_loan_application_v1.json.

Quick start

Prerequisites

• Python 3.12+ and Node 20+
• An Anthropic API key

Backend

cd backend
uv venv --python 3.12 .venv && source .venv/bin/activate   # or: python3.12 -m venv .venv
uv pip install -e ".[dev]"                                  # or: pip install -e ".[dev]"
cp ../.env.example ../.env       # add your ANTHROPIC_API_KEY
uvicorn app.main:app --reload --port 8000

API docs at http://localhost:8000/api/docs.

Frontend

cd frontend
npm install
npm run dev        # http://localhost:3000

Docker (both services)

cp .env.example .env   # add ANTHROPIC_API_KEY
docker compose up --build

Sample data

Three loan packages under sample-data/, surfaced as demo scenarios via GET /api/pipeline/scenarios:

Package	Expected outcome
acme_widgets	Clean approve — DSCR ~2.8x, LTV ~68%, growing revenue
bravo_logistics	approve_with_conditions — DSCR ~1.19x, declining revenue, elevated leverage
charlie_retail	Validation error path — only 2 of 3 required years present

Tests

cd backend
.venv/bin/python -m pytest                    # unit + route tests
.venv/bin/python -m pytest -m integration     # full agentic loop (needs ANTHROPIC_API_KEY)

Status notes

• The backend domain layer (tools, schema, prompt, agent loop, sample data, tests) is the workshop's teaching core and is fully implemented and tested here.
• The frontend streams the agent's reasoning and tool calls generically; its domain-specific visualization components are inherited from the upstream finance demo and are not yet retailored to loan underwriting (tracked as follow-up; not on the core path).