Fintech

−40% false positives on real-time fraud scoring

An LLM-augmented risk model scoring transactions in real time to catch fraud without blocking good customers.

The problem

A rules-based fraud engine had a high false-positive rate, blocking legitimate high-value transactions.

The solution

We combined a gradient-boosted baseline with an LLM reasoning layer over transaction narratives and device signals, scoring in under 200ms.

−40%

False positives

+22%

Fraud caught

<200ms

Scoring latency

Stack

XGBoostGPT-4o-miniKafkaFeature store