NFL QB Touchdown Predictor
Machine Learning System

Engineered a database-first ML workflow that ingests historical NFL data, constructs features, trains models, and exposes a Streamlit UI for “what-if” matchups.

• Designed SQLite schema and loaders for 1,536 games, ensuring referential integrity while supporting incremental updates.
• Engineered 42 features spanning QB form, defensive pressure, red-zone efficiency, and weather context; automated scaling/imputation pipelines.
• Benchmarked logistic regression vs. XGBoost with stratified CV; ensemble hit 88% accuracy and 0.91 ROC–AUC with calibrated probabilities.
• Wrapped the model in a Streamlit UI that surfaces SHAP values so coaches can see why predictions shift per matchup.

{
  "qb": "Patrick Mahomes",
  "opponent": "BUF",
  "pressure_rate": 21.5,
  "red_zone_trips": 4
}

{
  "will_throw_td": true,
  "probability": 0.87,
  "top_features": [
    "red_zone_trips",
    "opponent_pass_defense",
    "pressure_rate"
  ]
}