A real-time fraud detection engine that cut chargebacks by 57% in one quarter

Northbridge Payments processes millions of card-not-present transactions a month, and their rules-based fraud filter was both missing sophisticated fraud patterns and blocking legitimate customers. We designed a streaming data pipeline that scores every transaction against a trained model in under 80 milliseconds, combining device fingerprinting, behavioral signals, and historical patterns. The model runs alongside the existing rules engine during a shadow period so the team could validate accuracy before it started blocking traffic. Within the first full quarter after launch, chargebacks dropped 57% while false-positive declines on legitimate customers fell by nearly a third.