A regional GCC bank experiencing $12M in annual fraud losses deployed Anlage Digital's real-time ML fraud detection system — transforming from a rules-based overnight batch process to millisecond AI scoring at transaction time.
The client — a mid-size regional bank with operations in UAE, Bahrain, and Oman — was losing approximately $12M annually to payment fraud, primarily through card-not-present (CNP) fraud, account takeover, and money mule networks. Their existing fraud prevention relied on static rules maintained by a small team, which produced two significant problems: high false positive rates that blocked legitimate transactions (frustrating 1 in 25 genuine customers), and slow detection that caught fraud only in overnight batch runs — hours after losses had already occurred.
The bank needed a real-time ML system that could score every transaction at submission time, adapt continuously to new fraud patterns, and reduce customer friction from false positives — all while meeting Central Bank compliance requirements for model explainability.
Phase 1 — Data Foundation: We built a real-time feature store on AWS using Feast and DynamoDB, computing 200+ fraud-relevant features at transaction time: behavioral velocity, device fingerprinting, merchant pattern analysis, and cross-account relationship features derived from 3 years of transaction history.
Phase 2 — Model Development: We developed an ensemble model combining gradient boosting (XGBoost), graph neural networks (for detecting money mule rings), and a transformer-based sequence model for account takeover detection. Models were trained on 24 months of labeled historical data, with class imbalance handled using cost-sensitive learning.
Phase 3 — Real-Time Infrastructure: The scoring system was deployed on AWS SageMaker with sub-50ms p99 latency requirements, integrated with the bank's card authorization system via real-time API. MLOps pipelines on SageMaker Pipelines handle weekly automated retraining on new labeled data.
Phase 4 — Explainability & Compliance: SHAP-based explanations were implemented for every fraud decision, satisfying Central Bank requirements for model transparency and enabling fraud investigators to understand and act on model outputs.
Our BFSI AI team has deep experience in fraud detection, credit risk, and financial crime prevention. Let's discuss your specific fraud challenge.
Book a Discovery Call