Fraud Detection with LSTM Autoencoder
Real-time anomaly detection on financial transactions, deployed serverless on AWS Lambda.
Live financial transactions flow in as time-ordered events for real-time scoring.
Fraud is rare and ever-changing, so labelled data is scarce and rule-based systems miss novel patterns.
Detect anomalous transactions in real time with high recall and few false positives.
Engineered a deep LSTM autoencoder that learns normal transaction sequences; high reconstruction error flags anomalies. Containerised and deployed on AWS Lambda for live scoring, with a CI/CD pipeline for model versioning and automated redeployment.
Tuning the reconstruction-error threshold to balance recall against false positives on highly imbalanced data.
Improved fraud recall by 20% and reduced false positives by 35%.
Online learning to adapt thresholds as fraud patterns drift.