Cut fraud false-positives by 60% with a real-time ML scoring engine
Engagement
10 weeks
Team
1 ML lead, 2 ML engineers, 1 platform engineer, 1 designer
Industry
Financial Services
Stage
Series B
The numbers that mattered to the client.
What we walked into.
The client's existing fraud-detection system was rules-based — 700+ hand-written rules that flagged 12% of transactions for manual review. Of those, only 0.3% were actual fraud. The false-positive rate was destroying customer experience and overloading the 24-person fraud-ops team.
The client had tried to ship an ML model twice. Both attempts failed in production — the first because the model couldn't keep up with the transaction rate, the second because drift monitoring wasn't built and the model degraded silently over 6 months.
What we built.
We built a streaming fraud-scoring engine with a clear architecture: Kafka topic for transactions, consumer that calls the inference API, scoring result written back to Kafka for downstream action.
The model itself was a gradient-boosted-tree ensemble trained on 18 months of historical data. We avoided deep learning because the marginal precision gain (2%) didn't justify the 4x latency cost. The ensemble gives 94% precision at 180ms p99.
On the MLOps side, we built the full stack: feature store, model registry with code+data+metrics hash, drift monitoring on input distribution and prediction distribution, and a quarterly retraining pipeline with human approval for production rollout.
Tech stack
What changed for the client.
- False-positive rate cut 60% (from 12% to 4.8% of flagged transactions)
- Fraud-ops team headcount freed for higher-value investigation work
- Quarterly retraining keeps precision above 92% with no manual intervention
- Latency budget (200ms) never breached in 12 months of production
“They shipped our fraud-detection model in 10 weeks. Precision 94%, false-positive rate cut by 60%, ROI in 5 months. The MLOps discipline is what made it stick.”
Head of Risk
Series B Fintech
Want results like these?
Book a free 30-minute consultation. We'll review your situation and propose a small first milestone — same playbook as the case studies above.