Personal data product
NCAA 2026 Bracket Model
A data product built from public rankings, schedules, historical tournament data, MLX models, simulations, and readable bracket outputs.
Visit projectRole
Builder
Tags
Data ProductML WorkflowInteractive Showcase
Problem
I wanted a practical office-pool decision tool that went beyond gut picks by combining public basketball data, modeling, simulation, and a readable product interface.
Users
Bracket-pool participants and portfolio reviewers evaluating data sourcing, modeling, and product storytelling skill.
Ownership
- Owned the full workflow from data sourcing and cleanup through modeling, simulation, evaluation, and UI packaging.
- Built the static interactive showcase around the model outputs so the results were usable, not just notebook artifacts.
Hard Parts
- Collected and normalized data from NCAA NET, AP Top 25, Sports Reference, Bracket Matrix, official bracket, schedules, logs, boxscores, and historical tournament rows.
- Compared MLX logistic baselines, priors, a tree benchmark, and a production ensemble.
- Resolved the full field and turned probabilistic model outputs into readable bracket decisions.
Leadership
- Demonstrates the ability to make technical model work legible through product UI, narrative, and decision-support framing.
- Shows end-to-end ownership across data, modeling, evaluation, calibration, and frontend storytelling.
Shipped
- Office-pool decision tool and playful portfolio UI.
- 10,000 Monte Carlo simulations.
- Bracket paths, title odds, upset boards, model notes, and team search.
Impact
- Proves data sourcing, cleanup, feature engineering, modeling, evaluation, calibration, UI design, and storytelling in one visible artifact.
Signals
10,000 Monte Carlo simulations.68 of 68 field teams resolved.Production ensemble reached 68.5% validation accuracy.
Highlights
- Collected and normalized NCAA NET, AP, Sports Reference, Bracket Matrix, official bracket, schedule, and boxscore snapshots.
- Built historical training datasets, current matchup features, recent-form fields, model artifacts, bracket variants, and confidence reports.
- Trained and compared MLX logistic models, matchup features, a tree benchmark, and an ensemble path for calibrated picks.
- Packaged the result as a static interactive showcase with bracket paths, title odds, upset boards, model notes, and team search.
Tools
PythonPandasBeautifulSoupMLXModel evaluationMonte Carlo simulationStatic HTMLJavaScriptData visualization