SMOKE: NBA Shot-Making Metric — Applied Machine Learning & Sports Analytics
PythonGradient-Boosted TreesEmpirical BayesStatistical Validation
A transparent, validated measure of NBA shot-making: actual field goal percentage minus what a shot-difficulty model expected, with confidence intervals on every value.
- Led a four-person ASU capstone team building a shot-difficulty model over NBA tracking data, then continued the research as SMOKE (Shots Made Over Known Expectation), released as a fully reproducible public repository with a manuscript in preparation for the 2027 MIT Sloan Sports Analytics Conference.
- Modeled 128,069 tracked shots from 14 release-time features with gradient-boosted trees cross-fitted by game (held-out AUC 0.64 vs. 0.50 base rate); shrunk player estimates with DerSimonian-Laird empirical Bayes and published every value with bootstrap 95% intervals.
- Validated with eight pre-registered tests: at least as stable year over year as effective FG% (0.48 vs. 0.43), adds information about next-season shot-making beyond past efficiency (adjusted R² 0.09 to 0.16, p < 0.001), and penalizes no playing style; only 53 of 266 qualified players separate from league average in a season.
View project ↗
Multi-Architecture Generative Content Studio — Deep Learning Capstone
PyTorchTransformersDiffusionGANsCNNsRNNs
A unified generative system that produces story text, scene illustrations, character portraits, and background music from a single prompt.
- Built a unified generative system producing story text, scene illustrations, character portraits, and background music from a single prompt.
- Integrated GPT-2, Stable Diffusion, StyleGAN2/3, VGG16/19, and LSTM into one cohesive pipeline.
- Fine-tuned GPT-2 for narrative generation, entity extraction, and structured scene descriptions; implemented Stable Diffusion with LoRA + ControlNet for composition-controlled scenes.
View project ↗
Tempo — AI Scheduling Assistant
FastAPIFlutterPostgreSQLLLM / NLPConstraint Optimization
A cross-platform AI scheduling app that generates, compares, and applies alternative day/week plans.
- Built a cross-platform scheduling app (Flutter front end, FastAPI back end) that generates alternative day/week plans, compares them, and applies the one the user picks.
- Designed a constraint-based scheduling engine that orders tasks, builds the daily timeline, resolves conflicts, and merges blocks into a coherent schedule.
- Engineered a scenario pipeline supporting generation, diff-based comparison, trade-off analysis, and undo/redo state management.
PSID Panel Dataset Analysis — Longitudinal Economics Project
PythonPandasStatsModelsPanel Data
A longitudinal analysis of PSID microdata on U.S. income and wealth dynamics across five survey waves.
- Reconciled Panel Study of Income Dynamics microdata across 13 survey waves, mapping 117 wave-specific variable codes to 9 stable variables, and built an 18,740-row analysis panel covering five waves (1999 to 2015).
- Built the full pipeline: cleaning, reconciliation, regression modeling, and visualization, with Gini and top-decile inequality measures implemented from scratch.
- Found wealth roughly twice as unequal as income (wealth Gini ~0.85 vs income Gini ~0.45), with the median wealth-to-income ratio rising from 1.25 to 2.05 across the panel.
View project ↗
Scientific Visualization Pipeline (ParaView) — Self-Directed
ParaViewparaview.simpleVTKNumPy
A fully Python-scripted ParaView pipeline for 3D scientific data visualization.
- Built a NumPy/VTK dataset (3D Gaussian concentration field + analytic ABC-flow velocity) and a fully Python-scripted ParaView pipeline (paraview.simple).
- Generated volume renderings, Contour isosurfaces, slices, and Stream Tracer streamline tubes, with automated figure and orbit-animation export.
View project ↗