Real-time fraud triage at authorization speed
63%
reduction in false positives
14ms
p99 scoring latency at authorization
2.1×
more confirmed fraud caught per analyst-hour
The challenge
Where it started.
The client's rules-based fraud engine flagged 4% of all transactions for manual review — overwhelming analysts, delaying settlements, and still missing coordinated attack patterns that didn't match any single rule. Every false positive was a customer inconvenienced and an analyst-hour spent.
The approach
How we built it.
Deployed Vibalitics on the transaction stream with exactly-once ingestion at peak festival-season load.
Trained ensemble fraud models on three years of labeled disposition history, benchmarked against the incumbent rules engine.
Shipped explainable declines — every block carries feature attributions that survive chargeback disputes.
Zorat triage agents now assemble case files for the alerts that still need human eyes, cutting review time per case.