Fraud signals surfaced before losses hit the ledger
A payments processor needed anomaly detection that compliance could defend - not another black-box alert queue.
The challenge
Rule-based monitoring generated thousands of false positives. Analysts were drowning; leadership needed models that caught real fraud without slowing legitimate transactions.
How we approached it
Start with analyst workflows
Mapped how fraud teams triaged alerts today - then designed features and thresholds around those decisions.
Governed feature pipelines
Built streaming feature stores and batch enrichment so models trained on the same signals used in production.
Explainable scoring
Delivered anomaly scores with reason codes analysts could cite in escalation notes and regulator reviews.
Measure what matters
Tracked precision, recall, and analyst handling time - not just model accuracy on historical labels.
Outcomes
Fewer false positives
Alert volume dropped while confirmed fraud catch rates held steady across pilot corridors.
Faster triage
Analysts cleared high-confidence queues in half the time with clearer prioritisation.
Production rollout
Models moved from shadow mode to live scoring with compliance sign-off on monitoring and rollback.
Technologies involved
- Python
- Spark
- Kafka
- PostgreSQL
- scikit-learn
- MLflow
- Grafana
In their words
“We finally had anomaly detection our analysts trusted - and compliance could explain.”