Problem
A financial-services client needed real-time fraud detection on transaction streams — flag suspicious activity within milliseconds of a transaction, with low false positives that would otherwise burn customer-experience budget on legitimate users.
Approach
Real-time inference pipeline against a streaming transaction feed. Feature engineering combining transaction context (amount, merchant, geography, time-of-day) with behavioural signals (velocity, deviation from per-account baselines, peer-group comparisons). Gradient-boosted classifier for the primary scoring layer with explicit thresholds for hard-block, manual review, and pass. Continuous retraining as new fraud patterns emerged.
Stack
Streaming inference · gradient boosting · feature engineering · low-latency serving · monitoring
Outcome
Real-time fraud detection that meaningfully reduced losses across the engagement period, with the false-positive rate kept low enough that legitimate transaction throughput was not disrupted. Validated the pattern of "real-time inference at the transaction edge" rather than batch fraud review hours after the loss is already booked.