Real-Time Fraud Intelligence
A streaming ML system that scores transactions in-flight and explains every decision it makes.
Overview
Status: internal project. This page documents the engineering approach rather than a commercial outcome.
Fraud teams usually discover patterns after the fact. This system moves detection into the request path: features are computed from an event stream, a model scores each transaction, and the reasoning is written alongside the decision so analysts can audit it later.
Problem
Batch scoring is too slow when approval windows are measured in milliseconds, and black-box scores are impossible to defend during a dispute. The engineering constraint was latency, determinism and auditability at the same time.
Approach
Event-time windows feed a feature store that is read synchronously by the scoring service. Models are versioned, shadow-deployed and promoted only after offline and online metrics agree. Every decision stores the model version, feature snapshot and explanation vector.
Result
The result is a scoring path that behaves like production infrastructure rather than a notebook: observable, replayable and safe to roll back.