Fraud Detection Pipeline

Opening Scenario

You're on the payments platform team. Fraud losses jumped 40% last quarter; the existing system is a batch job that scores transactions hourly using a Python script, which means a fraudulent card is detected an hour after it's already been swiped at five merchants. The product ask: stop confirmed fraud at authorization time (sub-100ms decision), flag suspicious patterns for review within seconds, and feed labeled outcomes back into the model retraining loop. A vendor pitches their "AI fraud platform"; your CTO wants to know whether to build, buy, or hybrid. Before writing a design, you need to understand what a real-time fraud pipeline actually looks like, what each component contributes, and where the hard problems live.

---

Question Progression

1. Why is fraud detection a streaming problem, not a batch one?

Question. Fraud has always been detected by batch jobs. Why is streaming worth the operational cost?

Reasoning. Fraud…