Impact
After rollout we saw a significant drop in chargebacks and manual reviews on the segments covered by the engine. In practice that meant blocking or flagging hundreds of high-risk orders per month that previously would have gone to manual review or slipped through.
Performance
The requirement was that risk evaluation must not be a bottleneck to checkout, so we targeted under ~200 ms end-to-end per transaction for online scoring (including feature enrichment and external calls). Most requests stay well below that in production.
Architecture & Continuous Growth
Eval Risk is a large-scale digital product that combines REST APIs, Event-Driven Architecture, Microservices, and Real-Time Alerts into a cohesive fraud detection platform. The system is continuously evolving:
- REST APIs handle synchronous risk scoring on critical checkout paths
- Event-Driven backbone allows multiple consumers (rules engine, ML scoring, analytics, notifications) to react asynchronously
- Microservices architecture enables independent scaling and deployment of different components (feature enrichment, model inference, alert processing)
- Real-time alerting system notifies fraud analysts immediately when high-risk patterns are detected
This architecture allows us to add new fraud detection models, integrate external data sources, and evolve business rules without disrupting the core payment flow. The modular design means we can replay events for experiments, scale each service independently, and continuously improve detection accuracy based on new fraud patterns.