AI-Powered Fraud Detection Engine

The Problem

Legacy fraud detection relied on rule-based systems with ~85% accuracy and 2-3 hour detection latency. Rules-based systems generated 40% false positives, causing customer friction. Banks needed real-time detection with minimal false positives.

My Contribution

I led the requirements definition and stakeholder alignment for this fraud detection initiative. Working between compliance officers and the engineering team, I audited every existing rule set, interviewed AML analysts to map false-positive failure modes, and authored the functional specification that guided model development. I designed the UAT framework used to validate detection thresholds against live transaction samples and ran structured sessions with compliance leadership to define the precision/recall tradeoffs that balanced fraud capture with customer friction before cutover.

Process

1. 1

   Discovery

   Audited existing rule sets and interviewed compliance analysts to map false-positive failure modes.

2. 2

   Requirements

   Translated compliance requirements into functional specifications for the model development team, including precision/recall thresholds and explainability standards.

3. 3

   UAT Design

   Built a structured UAT framework using historical transaction samples to validate detection accuracy before shadow-run began.

4. 4

   Cutover

   Managed stakeholder sign-off across compliance, operations, and product during the 4-week parallel-run period before full cutover.

The Solution

Built a multi-model ensemble combining LSTM neural networks for behavioral patterns, XGBoost for transaction features, and graph neural networks for ring detection. Implemented streaming pipeline using Kafka for sub-100ms latency. Model retrained daily on new transaction patterns.

Results

• 99.2% detection accuracy (up from 85%)
• 87% reduction in false positives (40% → 5%)
• 98% of fraud caught within 2 minutes
• $12M+ in fraud prevented annually
• Scales to 10M+ transactions per day

Tech Stack