Background
A global fintech firm was facing operational bottlenecks in compliance due to high volumes of false positives in its transaction monitoring system. Manual triage of alerts slowed down response times to actual high-risk transactions, affecting regulatory compliance and lowering team productivity.
Solution
Affine developed a machine learning–based transaction screening system that augmented the client’s existing fuzzy match logic with predictive scoring. Key solution elements included: A hit-level ML model built using LightGBM, trained on historical payments and alert data Similarity score enrichment and advanced feature engineering leveraging data from payments, entities, banks, and case logs Decile-based probability scoring to flag high-likelihood true matches and suppress low-risk alerts Integrated MLOps pipelines for scalable model deployment, inference, and monitoring Explainable model outputs to support analyst decisions and fulfil compliance audit requirements
Impact
20% reduction in analyst workload by suppressing low-risk alerts $1 million annual savings in operational costs 15% faster resolution of high-risk transactions ~87% capture rate of true matches within the top deciles 89% AUC score, demonstrating strong model performance and reliability