FinTech
Automated risk assessment & fraud detection
Payments platform40M+ transactions / monthUS & EU
The problem
Fraud losses were growing faster than transaction volume. Risk models refreshed quarterly, so new typologies ran unchecked for months; manual review queues stretched to six hours per case while regulators pressed for explainable decisions. Every mitigation traded loss for friction, false positives were quietly taxing good customers.
What shipped
- Real-time scoring service, sub-50ms at p99
- Graph-based fraud feature store across accounts, devices and merchants
- Eval harness with golden fraud sets gating every model release
- Analyst case console, the model abstains, humans decide
Problem-to-production blueprint
1DiagnosisLoss decomposition across fraud typologies; the queue, not the model, was the first bottleneck.
2Architecture & PoCReal-time scoring plus graph feature store, proven against 90 days of labelled history.
3Production engineeringStreaming features, model gateway, eval harness, case console, hardened and load-tested.
4StewardshipWeekly drift reviews matured into fully automated daily refresh with rollback gates.
Outcomes
-42%False positives
11 minCase review, was 6 hrs
DailyModel refresh, was quarterly
Same problem, different book
Bring us the fraud and risk problem you have now.
We diagnose before we build, and one owner takes it from first decision to production.