Finance · Enterprise AI · Series C payments firm
Real-time fraud detection that decides in 8 milliseconds
We replaced a nightly rules batch with a streaming ML model that scores every transaction in real time — cutting fraud loss 41% while false positives fell.
PythonKafkapgvectorSageMaker
-41%
fraud loss
Timeline
4 months
Team
5 engineers + 2 data scientists
The challenge
What was breaking
- 01Fraud was reviewed in overnight batches, so losses were discovered too late.
- 02The rules engine produced 9% false positives, angering good customers.
- 03New attack patterns waited on a release cycle to be addressed.
The approach
How we tackled it
- 01Built a feature store from live transaction and device signals.
- 02Trained a gradient-boosted model with a calibrated threshold per segment.
- 03Served it on a streaming path with a hard 10ms budget and a human queue for the grey zone.
The solution
What we shipped
- 01Sub-10ms scoring on 100% of transactions.
- 02A self-tuning threshold that holds false positives under 1.5%.
- 03A feedback loop where every reviewed case retrains the model nightly.
In pictures
A closer look
The results
-41%
fraud loss
-83%
false positives
8ms
median decision
3x
faster to new patterns
"We went from chasing yesterday's fraud to stopping today's. The finance team finally trusts the green light."
Related work