A selection of things I've built and analyzed.
- 01 Research paper
Same Recipe, Different Answer
Base-Learner Sensitivity in Uplift Meta-Learner Benchmarks
A two-part empirical paper. Part 1 replicates Diemert et al. (2018) on the 14M-row Criteo Uplift dataset. Part 2 asks the question the field skips: what happens when you swap the base learner underneath each meta-learner recipe? The rankings break — Class-Transformation swings 82% in Qini just by changing engines (LR → XGBoost), and the S-learner on Hillstrom moves 5×. Only Causal Forest survives the swap. A study in which benchmark claims actually transfer — and which are artifacts of the model underneath.
Read the paper
- 02 Capstone · Deployed
Iqaluit Smart Water Delivery
Capstone with the City of Iqaluit — a mobile app + IoT + ARIMA forecasting stack to modernize trucked water delivery for 470+ households above the treeline. Real-time tank telemetry, driver routing, and demand prediction cut projected operational cost by $250K / year. Built the offline-first mobile UX, the tank-sensor backend, and the ARIMA model that anticipates when a house is about to run dry — before the driver has to guess.
Explore the project
- 03
Heart Disease Prediction
Built a supervised classifier on clinical patient data to predict heart-disease risk — feature engineering, model selection, and evaluation in Python. Compared tree-based models vs. logistic regression to isolate which clinical features actually move the prediction.
- 04
S&P 500 Stock Predictor
Time-series model that forecasts S&P 500 movements from historical price data — feature construction, backtesting, and error analysis. A study in how far pattern-based prediction can push against genuinely noisy financial signal.