Add REST API endpoint for real-time transaction scoring

[stabrea]

Summary

Wrap the trained fraud detection model in a FastAPI service that accepts transaction data and returns fraud probability scores in real time.

Motivation

• Production fraud detection requires low-latency scoring at transaction time
• REST API enables integration with payment processing pipelines
• Demonstrates deployment readiness beyond notebook-stage ML

Proposed Approach

1. FastAPI service:

   • POST /score endpoint accepting transaction JSON
   • Returns fraud probability, risk tier (low/medium/high), and top contributing features
   • Input validation with Pydantic models
   • Health check endpoint at GET /health

2. Model serving:

   • Load serialized model on startup
   • Feature preprocessing pipeline included in serving path
   • Model versioning support (load by version tag)

3. Containerization:

   • Dockerfile with multi-stage build
   • docker-compose.yml for local development
   • Environment-based configuration (model path, threshold, port)

4. Performance:

   • Target <50ms p95 latency for single transaction scoring
   • Add request logging for monitoring
   • Rate limiting for abuse prevention

Acceptance Criteria

• FastAPI app with /score and /health endpoints
• Pydantic request/response models with validation
• Dockerfile and docker-compose.yml
• Example curl commands in README
• Latency benchmarks documented