A three-tier fraud detection system. Cheap models clear the obvious cases; expensive models earn their keep on the ambiguous ones. Truly uncertain transactions get five specialized AI judges, who deliberate in turn and issue a reasoned verdict — then place a voice call and auto-draft the dispute paperwork.

Transaction
    ↓
┌──────────────────────────┐
│ Tier 1: XGBoost (CPU)    │ ~few ms
│   Kaggle 284K txns       │   Auto-clears the easy cases
│   ROC-AUC ~0.98          │
└─────────┬────────────────┘
          ↓ uncertain (0.30 < score < 0.95)
┌──────────────────────────┐
│ Tier 2: Sequence         │ ~tens of ms
│   Transformer (GPU)      │   Catches behavioural anomalies
│   6.5M params · AP ~0.83 │   across the last N transactions
└─────────┬────────────────┘
          ↓ still uncertain (0.20 < score < 0.85)
┌──────────────────────────┐
│ Tier 3: 5 AI Judges      │ mock: ~instant
│   (Pattern · Geography · │   real: ~5–10s with Qwen-72B
│    Temporal · Merchant · │
│    Chief)                │
└─────────┬────────────────┘
          ↓
   BLOCK + per-judge ElevenLabs voice alert + auto-dispute letter

backend/      FastAPI + the three tiers
frontend/     Next.js 15 landing pages + demo console
sampleui/     Source HTML/CSS the landing pages were ported from

cd backend
pip install -e .
# GPU box (Tier 2 training + optional Tier 3 real judges):
pip install -e ".[gpu]"

Grab the Kaggle creditcardfraud dataset and drop it at: backend/data/raw/creditcard.csv (~144 MB). Either drag-upload, or on the pod with kaggle CLI:

kaggle datasets download -d mlg-ulb/creditcardfraud -p data/raw --unzip

python -m ml.train               # Tier 1 XGBoost (CPU, ~30 s)
python -m ml.train_transformer   # Tier 2 Transformer (GPU, ~5–15 min on MI300X)

Mock mode (demo-safe, no GPU required for Tier 3):

SKIP_LLM=1 uvicorn main:app --host 0.0.0.0 --port 8000

Real Qwen-72B judges (requires a 192 GB GPU and the weights cached):

export HF_HOME=/workspace/.cache/huggingface          # writable cache
export ELEVENLABS_API_KEY=...                         # voice on every turn
export VOICE_CHIEF=... VOICE_PATTERN=... VOICE_GEOGRAPHY=...
export VOICE_TEMPORAL=... VOICE_MERCHANT=...
uvicorn main:app --host 0.0.0.0 --port 8000           # ~3 min cold load

If the real-LLM path errors mid-request (OOM, timeout, anything), the judge route silently falls back to the scripted mock dialogue so /analyze never 500s.

• GET /health — model load state
• GET /dataset/random?fraud=true|false — returns a Kaggle row + synthesized merchant/location + prior user-history for Tier 2
• POST /analyze — full pipeline: ML scores + judge dialogue + voice audio + dispute letter
• POST /ask-judge — follow-up question to a specific judge; returns voice audio

cloudflared tunnel --url http://localhost:8000
# or: ssh -R 80:localhost:8000 serveo.net

Then in frontend/.env.local:

BACKEND_URL=https://<your-tunnel>.trycloudflare.com

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

Routes:

• / — landing hero
• /mission — mission statement
• /technology — the 5-judge bento grid
• /demo — the real demo console (transaction picker → tier pipeline → court → ask-a-judge)

• Tier 2 on the Kaggle dataset is trained by clustering transactions into 2,000 pseudo-users (Kaggle has no user IDs) and feeding each new transaction as the last position of a length-50 sequence.
• Tier 3 mock dialogue uses ml_score to pick from fraud / legit fragment pools; confidence is derived from the ML tier scores, not hardcoded.
• /dataset/random?fraud=true pre-filters to rows XGBoost catches confidently, so the demo never gets a fraud sample the model silently approves.