Fraud Detection System
A classical machine learning system for detecting fraudulent transactions with a focus on precision/recall trade-offs.
Problem
Fraudulent transactions are rare and diverse, so naive classifiers either miss fraud or overwhelm review teams with false positives.
Solution
A classical ML pipeline explicitly tuned for the precision/recall trade-off inherent to imbalanced fraud data, rather than optimizing for raw accuracy.
Overview
Fraud is rare — often under 1% of transactions — which means a classifier optimized for accuracy alone can be 99% “accurate” while catching zero fraud. This project treats that trade-off as the core design problem, not an afterthought.
How it works
A classical ML pipeline is deliberately tuned against precision and recall rather than raw accuracy, with class-imbalance handling and a threshold chosen based on the real cost of a false negative (missed fraud) versus a false positive (an annoyed legitimate customer).
Result
A fraud flagging system whose evaluation metrics actually reflect what matters operationally, rather than a headline accuracy number that hides the real performance.
More projects
Product Recommendation Engine
An e-commerce recommendation engine combining a from-scratch ALS implementation (NumPy/SciPy) with content-based scikit-learn filtering.
Retail Analytics Warehouse
A dbt + DuckDB ELT pipeline over synthetic retail data, with 35 automated data-quality tests and a Streamlit dashboard.
Bike-Share Demand Forecasting
Hourly bike-share rental demand forecasting with honest time-series validation, SHAP explainability, and a containerized FastAPI serving layer.