What I'm Building.
My projects.
An end-to-end data platform forecasting NYC 311 service request demand across 6.2M records and 14 complaint categories. A production-style dbt + DuckDB pipeline feeds a hybrid Prophet/LightGBM forecasting model (27.6% mean MAPE) and a Gemini-powered natural language data assistant, all served through a Streamlit dashboard with live staffing recommendations.
View project
F1 pit strategy under safety-car uncertainty: a tyre-degradation and safety-car hazard model feed a CP-SAT solver that commits to a pit plan before knowing if a safety car comes out, then scores it against what actually happened and a perfect-foresight oracle. Across 652 driver-races, the strategy captured 58% of the value hindsight had available. Live solver studio built with FastAPI and React.
View project
A staggered-adoption causal study (Callaway-Sant'Anna) testing whether a country's internet actually gets faster once Starlink arrives, using Ookla's public speed-test data rather than the marketing. The naive comparison came back negative; correcting for a confounded control group revealed a real but modest effect, a few Mbps rather than the dramatic jump the hype implies.
View project
An interactive Tableau dashboard that helps you find where your salary goes furthest based on what matters most to you. A Python pipeline combines data from five sources across 618 cities to power an interactive world map, personalised rankings, city comparisons, and a supporting data story.
View project
Predicts what a specific NFL head coach will actually do on 4th down, learned from their real decision history rather than the standard win-probability-optimal recommendation. An XGBoost model and a PyTorch per-coach embedding model are ensembled on nflverse play-by-play data. Served via FastAPI with a weekly automated retrain, behind an Astro frontend.
View project
Aegis Run
A full-stack fantasy draft game built on a game-agnostic simulation engine: draft a roster, then simulate a group stage and bracket run against a team-strength and synergy model built from scratch. Includes a deterministic daily challenge with four independently-varying rulesets, and a Postgres/Drizzle data layer deployed on Vercel from an npm-workspace monorepo.
View project
A publicly scored earthquake rate forecast for New Zealand: forecasts are timestamped and committed before each window closes, then scored in the open against the same statistical tests seismologists use. Most forecasting tools get evaluated on data they've already seen; this one commits before the window closes, so every prediction is genuinely testable. Live and accumulating; results are still pending.
View project