Project arc
Built for an office March Madness pool, packaged like a product demo.
The goal was practical: make sharper bracket picks for a friendly office pool without relying on gut feel alone. The portfolio version shows the full path from messy public basketball data to a clean, explorable decision tool.
Pick a bracket that balances chalk, title upside, and selective upset risk.
Readable outputs first, model details available for anyone who wants to inspect the logic.
Research, scraping, cleaning, modeling, simulation, UI design, and storytelling in one piece.
NCAA NET, AP, Sports Reference, Bracket Matrix, official bracket, schedules, and boxscore snapshots.
Normalized team names, resolved source mismatches, merged field context, and tracked fallback quality.
Recent-form windows, power ratings, seed priors, efficiency edges, pace, shooting, rebounding, and turnover matchups.
Built tournament matchup rows, trained MLX logistic models, and compared a tree-based benchmark.
Blended matchup model, tree benchmark, heuristic strength, and historical seed priors for calibrated picks.
Packaged office-pool picks, round-reach odds, upset risk, data quality, and team search into a static UI.