[FEATURE] Fraud Detection API — Serve trained Deep SVDD + GraphSAGE models for real-time scoring

[gelluisaac]

Description

Expose the trained anomaly detection models (DeepSVDD + InductiveGraphSAGE) through
a REST API so the web dashboard can display real fraud predictions.

Endpoints

• POST /api/v1/fraud/score — Score one or more accounts
  • Input: {"accounts": ["GABC...", "GDEF..."], "edges": [...]}
  • Output: {"scores": {"GABC...": 0.85, "GDEF...": 0.12}}
• GET /api/v1/fraud/alerts — Recent fraud alerts (paginated, filterable by risk level)
• GET /api/v1/fraud/stats — Aggregated fraud statistics (total alerts, high/med/low, risk over time)

Implementation Details

• Load model checkpoints from the model registry on startup
• Implement a model caching layer to avoid reloading on every request
• Handle the case where models are not yet trained (return 503 or graceful message)
• Input validation via Pydantic schemas

Acceptance Criteria

• /fraud/score returns scores in <500ms for batches up to 50 accounts
• /fraud/alerts returns data matching FraudStats type in the frontend
• Model loading is cached and doesn't block the server startup if models are absent
• Graceful error handling for missing or corrupted model checkpoints

Dependencies

• Issue Build Horizon Streaming Client #2 (Database models) for storing alerts
• Existing astroml/pipeline/scoring.py and astroml/models/deep_svdd.py

Labels

enhancement, api, ml, fraud-detection