Research work on mechanistic interpretability of neural networks for AI safety

• Bootstrapped the company’s machine learning system for online payment fraud detection.• Implemented targeted solutions for credit card fraud, account take over, promo abuse.• Introduced better error metrics for model diagnosis under very high class skew.• Wrote the decision engine for model-informed as well as rule-based real time decisions.• Developed data pipelines for online and offline feature generation and storage.• Built a control-group based simulation framework to accurately measure model performance.

• Investigated payment fraud trends by identifying anomalous events in financial transaction data.• Optimized decision thresholds to reduce the Risk system's overall cost of misclassification.• Researched the information value of new variables to include in machine learning models.

• Wrote software to clean, combine and ingest data from various external sources.• Evaluated the effectiveness of classifier algorithms via AUC-ROC analysis

• Software development • Query optimization • Systems Integration • Incident management