Real-Time Fraud Detection

Americana Foods — operator of major food brands across the region — was losing revenue to point-of-sale fraud that manual review processes caught too slowly, if at all. Transaction logs and surveillance footage existed in separate systems, so investigators had to reconcile them by hand.

I developed the fraud detection system as a junior developer: the computer-vision pipeline, the API integrations connecting transaction data to surveillance feeds, and the matching algorithms between the two.

The core idea was correlation: a transaction record alone can look legitimate, and footage alone is just video — but mapped together with timestamps and register positions, anomalies become visible. I built OpenCV-based processing over surveillance streams and engineered algorithms to parse and map transaction data onto them, flagging mismatches in real time rather than in post-hoc review.

The system identified and blocked 95% of fraudulent transactions in real time across 50+ POS endpoints during production deployment. Fraud traceability improved by 30%, and investigation review time dropped by 40% because evidence arrived pre-correlated.

The most valuable ML systems often aren't novel models — they're well-engineered joins between data sources nobody had connected. This project also sharpened my instinct for responsible AI: when software blocks transactions in real time, the cost of a false positive is a human problem, not just an engineering one.