Guardrail — AI Hallucination Audit Suite

A professional hallucination detection API for AI responses.

Architecture (Cascading Audit Pipeline)

1. Extractor (SpaCy) — Named entity mismatch detection (local, ~10ms)
2. Scorer (Cross-Encoder via ONNX) — Semantic faithfulness score (local, ~10ms)
3. Judge (Groq Llama 3.3 70B) — Natural language explanation (API, ~1s)
4. Memory (Supabase) — Audit log persistence + leaderboard

Stack

• Backend: FastAPI + Python 3.13
• NLP: SpaCy (NER) + fastembed (ONNX cross-encoder, no PyTorch)
• LLM: Groq Llama-3.3-70b-versatile (zero cost, sub-second)
• Database: Supabase (Postgres)
• Hosting: Render (free tier — 512 MB RAM)

Endpoints

Method	Path	Description
GET	/health	Service health check
POST	/audit	Run full hallucination audit
GET	/leaderboard	Aggregated model scores

Local Development

python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_sm
cp .env.example .env  # fill in your keys
uvicorn main:app --reload

Open http://localhost:8000/docs for the interactive API docs.

License

MIT