Fraud Detection Microservice

Problem statement

Card present and e-commerce traffic introduced delayed chargebacks and manual review bottlenecks. The existing rules engine could not score in under 150ms end-to-end, and it missed coordinated fraud patterns across merchants.

Architecture overview

The service sits on the critical path of authorization. A Kafka consumer hydrates features from Redis, applies deterministic rules, and calls a lightweight model hosted in a separate pod. Scores are written back to the transaction stream for downstream review tooling.

Technical decisions & tradeoffs

• Chose Redis for sub-millisecond feature retrieval, accepting eventual consistency for non-critical features.
• Split model inference into a separate deployment to keep the scoring service lean and horizontally scalable.
• Used idempotent message handling to tolerate duplicates from the stream at the cost of extra storage.

Lessons learned

Latency budgets are strict in payment flows. Keeping inference isolated made it easier to tune performance without touching transaction processing.