Production ML System
Fraud Risk Ops Platform
Model serving · policy decisions · audit trails · operations.
Turns model scores into policy-driven decisions with stable APIs, review workflows, audit trails, background jobs, and monitoring.
Problem
Fraud scoring only becomes useful when analysts can act on model output through clear thresholds, review queues, audit trails, and monitoring signals.
Approach
- Separates model scoring, policy decisions, validation, audit trails, persistence, batch jobs, and monitoring so each layer can be reviewed independently.
- Uses configurable threshold policies to turn calibrated risk scores into approve, decline, or review actions without changing the model artifact.
- Pairs the FastAPI inference service with an analyst-facing Streamlit console for score review, diagnostics, and operational decision support.
Design Decisions & Trade-offs
- Model scoring is separated from policy decisions so thresholds and review behavior can evolve without changing the model artifact itself.
- Readiness fails closed when artifacts, schemas, checksums, policies, or runtime compatibility are invalid rather than silently substituting heuristic behavior.
- SQLite remains the operational source of truth for jobs and audit records, while Redis is used only as a wake-up queue, keeping the local runtime durable and inspectable without unnecessary queue infrastructure.
Signals & Results
- Operational boundary: model score → policy decision → audit trail → review workflow → monitoring.
- Engineering evidence: versioned API contracts, readiness checks, persisted jobs/audit records, Redis-backed worker flow, Prometheus metrics, Grafana provisioning, Docker, and CI.