• Built and evaluated machine learning models to predict U.S. residential property prices using large-scale listing data from Cotality’s Trestle API.
• Owned an end-to-end pipeline (EDA → Feature Engineering → Model Tuning → Packaging).
• Optimized Ridge/Random Forest/XGBoost; reached strong fit and stable error (MDAPE focus), achieving R² ≈ 0.89 and MDAPE ≈ 6% on held-out data.
• Produced reproducible Jupyter workflows and model diagnostics to identify major drivers of housing valuation and support iterative model improvement.

GitHub: Home Price Predictions

• Developed ARIMA and decision-tree forecasting models that improved demand forecast accuracy by 15% and reduced inventory stock-outs by 20%.
• Built Tableau KPI dashboards used to monitor supply-chain performance and support faster operational decision-making.
• Analyzed sales and inventory patterns to improve planning accuracy and reduce avoidable supply disruptions.

• Built ARIMA-based forecasting models that reduced material procurement costs by 15%, improved planning accuracy by 10%.
• Digitized expense and procurement tracking workflows, improving financial visibility and supporting more accurate budgeting.
• Combined historical sales and external market indicators to strengthen demand planning and purchasing decisions.

• Built a full ML pipeline to predict U.S. home prices using 80k+ housing records (CRMLS & Trestle API).
• Cleaned, encoded, and engineered features (log transforms, interaction terms) for model stability.
• Tuned Ridge & Random Forest models achieving R² = 0.89 and MDAPE ≈ 6 % with diagnostics used to explain the strongest pricing drivers
• Visualized key drivers using feature importances and residual diagnostics.

• Analyzed 1M + Food.com ratings to test if unhealthy recipes receive higher scores.
• Engineered a GridSearch-tuned scikit-learn pipeline (ColumnTransformer → RandomForest) with SHAP explainability.
• Published an interactive HTML report visualizing rating bias patterns and model insights.

• Trained Random Forest and Deep-Learning (TensorFlow) models for charge-off prediction.
• Achieved 94.9 % accuracy (↑ 6 pp over logistic baseline).
• Delivered interactive notebooks for threshold tuning and misclassification review.

• Used GeoPandas + PySAL to analyze 1M + NYC/Chicago records for station-adjacent crime risk.
• Applied GWR and clustering to map crime hotspots and socioeconomic disparities.
• Produced interactive maps supporting equitable city-planning insights.

📄 View Credential (PDF)