Productionizing a Fraud Model: Rollout, Monitoring and a 50 ms p99 Budget

Reported in interviews at Capital One

Problem statement

How would you put a card-fraud model into production? A tree ensemble scores a heavy stream of authorizations, and decisions must be almost instant. High scores trigger declines or review, but chargeback labels take months and analysts resolve only a sample sooner.

Explain safe deployment, then diagnose a feature failure and a latency bottleneck.

Show hint

Which parts of this system change slowly and which change fast, and how does the months-long label delay limit what you can measure at launch?

Relevant roles: ML Engineer and MLOps Engineer.

Related free problems