Fintech ML · 2024
Realtime Fraud Scoring
A low-latency scoring API serving fraud decisions in under 15 ms, monitored end to end.
Realtime Fraud Scoring
- Role
- ML engineer
- Timeline
- 6 months
- Team
- 2 ML engineers, 2 backend engineers, 1 risk analyst
- Focus
- PyTorch, FastAPI, Kubernetes
The problem
Rule-based fraud checks missed new attack patterns and blocked too many genuine customers. The business needed a model that could score every transaction in real time.
My role
I designed the model and the serving architecture, built the scoring API, and set up monitoring and a safe rollout with the backend team.
Key decisions
- Latency as a requirement. A compact model and a feature cache kept scoring under 15 ms at the 99th percentile.
- Shadow mode before launch. The model scored live traffic silently for weeks so we could compare it to the rules before it made decisions.
- Monitor the data, not just the model. Input drift and score distribution alerts caught problems before they became incidents.
Outcome
Fraud caught increased while false declines dropped, and the API has met its latency target since launch. Next time I would invest in automated retraining sooner.