🏥 AI-Powered Medical Billing Platform

Claim Cruncher

A production-ready medical billing intake platform with Claude AI, streaming intelligence, Tesseract OCR, HIPAA-ready architecture, and real-time denial analysis. Built from scratch in 11 days.

🎁 A wedding gift · built by Shane · for Gavin Brazelton · May 2, 2026

⬡ View on GitHub 🚀 Quick Start 📖 What You Built

~5k

Lines of Code

New Files

Claude AI Modes

DB Migrations

Days Built

✦

Cruncher AI

Claude claude-sonnet-4-6 · Claim #CLM-2024-0847 loaded

● Streaming

What CPT codes are billed on this claim? Any potential issues?

✦

🔧 get_claim 🔧 get_claim_lines

This claim bills 99213 (Office visit, est. patient, moderate complexity) and 93000 (ECG with interpretation). A few flags worth noting:

⚠️ Modifier 25 is missing on 99213 — required when billed same-day as a procedure. Payer may bundle and deny the E&M.

⚠️ ICD-10 I10 (HTN) doesn't directly support 93000 — consider adding R00.1 (bradycardia) or Z82.49 (family cardiac history).

This claim was denied. Generate an appeal letter.

✦

🔧 flag_claim 🔧 search_similar_claims

Ask about this claim…

→

What You Built

A Real Company's Software Stack

This isn't a tutorial project. It's a full production architecture — the same stack that medical billing SaaS companies charge $300/month per seat to run.

🏗️

FastAPI Backend

Async Python API with SQLAlchemy ORM, JWT auth, role-based access control (6 roles), HIPAA audit middleware that logs every request, and soft deletes throughout.

FastAPISQLAlchemyJWTRBACAsync

🤖

Claude AI — 4 Modes

Streaming agentic chat with tool use (Claude calls your DB live). Auto-flag billing errors with Haiku. Denial analysis with dispute likelihood scores. EOB document parsing.

claude-sonnet-4-6Tool UseSSE StreamHIPAA Guard

🔍

Tesseract OCR Pipeline

Upload a PDF or image → Tesseract extracts text at 300 DPI with per-page confidence scoring → structured JSON via Claude → auto-flag billing issues → RAG indexed. All async via arq queue.

Tesseractpdf2imagearqRedis

🧠

pgvector RAG

Every claim is vectorized and stored in PostgreSQL with the pgvector extension. Claude can search semantically similar historical claims to surface patterns and catch repeat denials.

pgvectorEmbeddingsCosine SearchLocal 128-dim

⚛️

React Frontend

TypeScript + Vite + Tailwind. Dashboard, claims list, claim detail with Cruncher AI panel, patients, facilities, tickets, credentials, reports. Full auth flow with token refresh.

ReactTypeScriptViteTailwind

🔒

HIPAA-Ready Architecture

PHI de-identification regex strips SSN, phone, email before every Claude API call. Audit middleware logs every request. Soft deletes preserve data lineage. BAA flag bypasses guard if signed.

PHI GuardAudit LogSoft DeleteBAA Flag

Architecture

5-Layer Production Stack

Every layer chosen for real-world scale — not just to make it work, but to make it grow.

Frontend

React + Vite

TypeScript, Tailwind, SSE streaming, token refresh interceptor

API

FastAPI

Async Python, 10 routers, RBAC, audit middleware, JWT

AI Layer

Claude API

Tool use agentic loop, 4 system prompts, streaming SSE

Database

PostgreSQL

14 migrations, pgvector extension, ivfflat ANN index

Worker

arq + Redis

Async job queue, Tesseract OCR, cloud fallback hooks

Quick Start

Running in Under 10 Minutes

Full stack — database, API, worker, and frontend — with one terminal each. You need Docker, Node 20, and Python 3.11.

Clone and configure

Copy the env template and drop in your Claude API key. Get one free at console.anthropic.com.

bash

# Clone the repo

$git clone https://github.com/thebardchat/Greenfield.git && cd Greenfield

$cp .env.example .env

# Open .env and set these two values:

ANTHROPIC_API_KEY=sk-ant-api03-...

JWT_SECRET=some-long-random-string-here

Start infrastructure

Docker Compose spins up PostgreSQL (with pgvector), Redis, and MinIO. Wait ~10 seconds for Postgres to initialize.

bash

$docker-compose up -d postgres redis minio

✔ Container greenfield-postgres-1 Started

✔ Container greenfield-redis-1 Started

✔ Container greenfield-minio-1 Started

Run migrations + seed

14 SQL migrations build the full schema including pgvector embeddings table. The seed loads demo orgs, users, and claims.

bash

# Run all migrations

$for f in db/migrations/*.sql; do docker-compose exec -T postgres psql -U claimcruncher -d claimcruncher < "$f"; done

CREATE TABLE ... CREATE INDEX ... (×14)

$python db/seeds/hash_passwords.py && docker-compose exec -T postgres psql -U claimcruncher -d claimcruncher < db/seeds/seed.sql

✓ Demo data seeded — login: admin@claimcruncher.com / Admin1234!

Start the API, worker, and frontend

Three terminals. Once all three are running, Claim Cruncher is live.

Terminal 1 — API

$cd apps/api && pip install -e ".[dev]"

$uvicorn app.main:app --reload --port 8000

INFO: Uvicorn running on http://0.0.0.0:8000

Terminal 2 — OCR Worker

$cd apps/worker && pip install -e ".[dev]"

$arq app.main.WorkerSettings

Worker started. Redis: redis://localhost:6380/0

Terminal 3 — Frontend

$cd apps/web && npm install && npm run dev

➜ Local: http://localhost:5173/

Verify Cruncher AI is connected

One curl to confirm Claude is wired up and streaming is ready.

bash

$curl -s http://localhost:8000/api/cruncher/health | python -m json.tool

{

"status": "ok",

"cruncher_enabled": true,

"model": "claude-sonnet-4-6",

"flag_model": "claude-haiku-4-5-20251001",

"rag_enabled": true

}

✓ Open http://localhost:5173 → log in → click any claim → Cruncher AI

API Reference

REST Endpoints

Full OpenAPI docs at http://localhost:8000/docs when running locally.

Method	Path	Description
POST	/api/auth/login	Login, returns JWT access + refresh tokens
GET	/api/claims	List claims (org-scoped, paginated, filterable)
POST	/api/claims	Create new claim with patient + payer info
GET	/api/claims/{id}	Get full claim with lines, documents, tickets
POST	/api/documents/upload	Upload PDF/image → SHA-256 dedup → OCR queue
GET	/api/documents/{id}/ocr	Get OCR status, extracted text, structured JSON
POST	/api/cruncher/chat	Streaming SSE — Claude tool-use agentic chat ✦ AI
POST	/api/cruncher/analyze-claim/{id}	Auto-flag billing errors with Haiku ✦ AI
POST	/api/cruncher/denial-analysis/{id}	Denial reason analysis + appeal letter language ✦ AI
POST	/api/cruncher/parse-eob	Parse EOB document text → structured JSON ✦ AI
GET	/api/reports/claims-summary	Status breakdown, collection rate, denial rate
GET	/api/reports/denial-trends	Top denial reasons + appeal success rates
GET	/api/reports/export	Streaming CSV export with full filtering
GET	/api/cruncher/health	Check Claude connection, models, RAG status

Technology

The Stack

🐍

Python 3.11

Backend runtime

⚡

FastAPI

Async API framework

⚛️

React + TypeScript

Frontend

🗃️

PostgreSQL

Primary database

🧩

pgvector

Semantic search

🤖

Claude API

AI intelligence

🔎

Tesseract OCR

Document scanning

📮

arq + Redis

Async job queue

🗄️

MinIO

S3-compatible storage

🐋

Docker Compose

Infrastructure

⚗️

Vite + Tailwind

Frontend build

🔐

JWT + RBAC

Auth + permissions

🎁

For Gavin

This is a real business you can launch. Medical billing offices pay $200–500/month per seat for software with a fraction of these features. You have HIPAA-ready architecture, Claude AI baked in, Tesseract OCR, and a complete React frontend — all built and waiting for you.

The only thing missing is your first customer.

Built with love by Shane Brazelton · Hazel Green, Alabama
Wedding gift · May 2, 2026 · "Best thing I've ever built you."