Medical AI content integrity monitor for the Google Cloud Rapid Agent Hackathon - Arize track.

Paste AI-generated health content and get a PASS / FAIL verdict with full evidence. MedCred combines OpenFDA checks, external API verification, Phoenix trace-informed judging, and calibration-gated self-improvement.

Calibration: 99% label agreement on a 100-case held-out set (40 PASS / 40 FAIL / 20 edge). Reproduce with POST /calibration/run?mode=fast.

OpenFDA extraction feeds the Gemini judge; a deterministic safety gate enforces FAIL after the verdict for invented drugs, deprecated treatments, and high-risk patterns. Phoenix handles subtler fraud the gate cannot catch (fake PMIDs, inflated efficacy claims).

Grounding → External APIs → OpenFDA → Phoenix MCP → phoenix_context → Judge → Gate → Annotate → Correction

Every 5 evaluations: read FAIL spans → propose prompt update → A/B test on held-out calibration set → promote only if agreement meets or exceeds baseline.

• Python 3.13+, uv
• Node.js 18+ (frontend)
• GCP project with Vertex AI + Agent Builder (optional grounding)
• Arize Phoenix API key

git clone https://github.com/shaileshdev4/medcred
cd medcred/backend
cp .env.example .env
# Set: PHOENIX_API_KEY, PHOENIX_SPACE, GOOGLE_CLOUD_PROJECT, GOOGLE_GENAI_USE_VERTEXAI=1
uv pip install -r requirements.txt
uv run python main.py

Health check: curl http://localhost:8000/health

cd medcred/frontend
cp .env.example .env.local
# Set NEXT_PUBLIC_API_URL=http://localhost:8000
npm install && npm run dev

Open http://localhost:3000

cd medcred/backend
uv pip install -r requirements-dev.txt
uv run pytest tests/ -q

Preflight (DNS + Phoenix + Vertex):

uv run python agent.py --preflight-only

ADK batch smoke (1 article):

uv run python agent.py healthgpt --smoke

Phoenix OTLP smoke:

uv run python scripts/verify_phoenix_trace.py

Calibration (100 cases, ~25 min; batch if you hit Vertex 429):

uv run python scripts/run_calibration_report.py
# or: curl -X POST "http://localhost:8000/calibration/run?mode=fast"

Pre-seed Phoenix traces before demo (25 evals):

uv run python scripts/preseed_phoenix_traces.py

medcred/
├── backend/
│   ├── agent.py              # ADK SequentialAgent + CLI
│   ├── main.py               # FastAPI API
│   ├── core/                 # Phoenix tracing, evaluator base, self-improve
│   ├── single_eval.py        # Analyze pipeline
│   ├── sse_handler.py        # Theater SSE
│   ├── phoenix_mcp.py        # ADK McpToolset → @arizeai/phoenix-mcp
│   ├── phoenix_context.py    # Drug-aware trace injection
│   ├── calibration_runner.py # Label agreement + promotion gate
│   ├── evaluators/           # Gemini medical judge
│   ├── calibration_dataset.json
│   ├── scripts/              # Smoke, calibration, preseed
│   └── tests/                # 53 tests
├── frontend/                 # Next.js 14 dashboard
├── images/                   # Architecture diagrams
├── docs/
│   ├── SUBMISSION.md         # Deploy + Devpost checklist
│   └── AGENT_BUILDER_MEDCRED.md
├── medcred_devpost.md        # Devpost write-up
└── LICENSE

See docs/SUBMISSION.md for Railway (backend), Vercel (frontend), and Neon (Postgres).

Docker:

docker build -t medcred-backend ./backend
docker build -t medcred-frontend ./frontend

Python, FastAPI, Next.js, Google ADK, Gemini 2.5 Flash, Vertex AI, Arize Phoenix, phoenix-mcp, OpenFDA, DailyMed, PubMed, ClinicalTrials.gov, MedlinePlus, PostgreSQL, Neon.

Full Devpost narrative: medcred_devpost.md

MIT - see LICENSE.