Real-Time Fraud Detection for a Regional Digital Bank
Cliente: Northbridge Digital Bank
Replaced a static rules engine with an adaptive fraud detection system that cut false positives by more than half while catching more genuine fraud.
-58%
False positive rate
+23%
Fraud caught
-31%
Customer card declines
El Desafío
Northbridge's existing rules-based fraud system had accumulated years of manually tuned thresholds, and by the time we engaged, it was generating so many false positives that genuinely fraudulent transactions were getting lost in the noise its own alerts created. Fraud analysts were triaging hundreds of flagged transactions a day with no good way to prioritize the ones that actually mattered, and legitimate customers were having cards declined at a rate that was visibly hurting retention. Any replacement needed to preserve full audit visibility for the compliance team — a black-box model that couldn't explain its decisions to a regulator was not an option.
La Solución
We deployed a model that scores transactions in real time using behavioral and device signals layered on top of standard transaction data, with a continuous feedback loop from confirmed fraud cases that keeps the model current as fraud patterns shift. Rather than replacing the bank's existing rules engine outright, the model runs as a scoring layer ahead of it — every decision the model makes is logged with the specific signals that drove the score, giving the compliance team the same audit trail they had before, just attached to a system that's dramatically more accurate. Analysts now see a ranked queue instead of an undifferentiated alert flood, and the false-positive rate dropped enough that legitimate customers stopped noticing card declines almost entirely.