Technical deep-dives into systems I've built. Each project solves a specific problem.
Research & Quantum
Merit scholarships require keeping a GPA, and STEM grades are the hardest to keep. Across 570K records, strong merit-aid exposure cut STEM completion by 0.7-1.3 points. A Stata replication of Sjoquist & Winters.
ML & AI
My CPI model scored RMSE 0.39, 4th of 50+ hackathon teams. Then I audited it: a one-line random walk scores 0.363. This is the autopsy of my own leaked metric and an honest rebuild of the backtest.
Data & Visualization
The obvious pacing benchmark flagged 77% of a decade of ticket history as behind. I redesigned it to exact calibration and shipped a live dashboard over 119K transactions and $11.4M in sales.
A brewery spent 25% of sales visits on accounts with no room to grow while 30% with real potential went under-visited. K-means found the mismatch; a genetic algorithm now routes each day's deliveries.
An agent-based model of the 1914 Ludlow tent colony, adapted from NetLogo's Fire model. Two sliders, tent density and company pressure, flip destruction from under 1% of the colony to 97%. Runs in your browser.
After the earthquake, a plane landed at Port-au-Prince every 17.5 minutes and supplies still fell short. Our 45-day Simio simulation showed staged capacity additions lifting delivered water by 67%.
Systems & Infrastructure
Much of the work in this project sat around the model, not in it: R and Python in one Quarto document, packages managed with renv, GitHub Actions rendering the documents automatically, and a full rebuild after my laptop was reimaged. My first API answered with two words: Hello, Swayam.