Built a real-time ML risk scoring engine that cut fraudulent transactions by 42% while dropping false positives to under 2%.
A growing fintech platform was relying on a rule-based risk engine that generated too many false positives — blocking legitimate transactions and frustrating users. The team needed a smarter, lower-latency solution without rebuilding the entire stack.
We designed and built a real-time ML risk scoring service that evaluates every transaction in under 10ms using a gradient-boosted ensemble model trained on historical behavioral data.
Transaction events are ingested via Apache Kafka, inference runs in a low-latency Python service backed by TorchServe, and every decision is written to PostgreSQL with a full audit trail.
