Things I've built

Processes confocal PL maps of hBN before and after irradiation: crops to the real scan area, fits and clusters bright spots, and compares conditions across 24 datasets from 3 instruments. An early version quietly distorted the data because I had tuned it to look right before I understood the physics well enough to judge it. The rebuilt version is the one I use now.

Parses SRIM/TRIM output for He⁺ and H⁺ ions in thin hBN on SiO₂: boron and nitrogen vacancy depth profiles, vacancies per ion, substrate contamination, and displacements per atom versus fluence. The output is a table of fluences that hit a target VB⁻ density without overdamaging the film.

Given an ion, energy, and fluence, labels the expected damage regime in hBN (below threshold, sparse, dense, overdamaged) and returns the closest experiments from a hand-built table of about 20 published irradiation studies.

Two desktop tools for the bench. One overlays PL maps on optical images and computes per-flake PL from hand-drawn regions. The other reads NT-MDT .mdt AFM files directly (no Gwyddion), levels them, measures flake thickness, and computes ISO 25178 roughness. Both feed one per-flake spreadsheet.

Team QAT's entry: does adding quantum-derived features (IQP encodings, quantum kernels) improve return prediction? I led the classical pipeline: 39 engineered features, Ridge/Lasso/gradient boosting baselines, then fusing classical and quantum signals, which moved MSE from 4.55 to 2.69 and IC from 0.932 to 0.961.

A pairs-trading backtester written to learn the mechanics firsthand. Finds cointegrated pairs, generates z-score signals, and sweeps 50+ configurations in parallel with transaction costs and walk-forward optimization, reporting PnL, Sharpe, and max drawdown.

Options pricing with full Greeks, portfolio analysis, and volatility forecasting in one Streamlit app, built for working through the math rather than trading.

NBA analytics over 1.09M shot locations and 87.5k five-man lineups across five seasons: shot-quality metrics, PCA player embeddings clustered into roles, and a gradient-boosted lineup model (R² = 0.71).

A three-person entry in a trading-bot competition on a simulated exchange. My bots: an Avellaneda–Stoikov fair-pricing market maker, plus mean-reversion, arbitrage, signal-driven, and prediction-market strategies. The backtester was shared work.

An AR whiteboard built with Vishal Nagamalla: notes anchored to real surfaces and shared live. I wrote the camera tracking and most of the capture app. Keeping the board stable while the camera moves was the core problem, and it took several rewrites of the homography and optical-flow tracking before it held up outside controlled conditions. Vishal built most of the Gemini service that answers questions about the board.

A mental-health chatbot for adolescents, serving 6,000+ monthly users across five schools. I co-founded it, trained the TensorFlow model, and built the React front end. Most of the work went into making responses useful rather than merely on-topic.

A double-entry payment ledger in Go: balanced transfers, idempotent writes, and balances that stay correct under concurrent transfers, backed by PostgreSQL with CI.

Connects Claude to the Onshape CAD API so a model can read documents, parts, and assemblies. The goal is generating VEX parts from a plain-language description.

A camera-first calorie tracker that installs to a phone as a PWA. Reads nutrition labels on-device, searches 7,277 foods offline, and syncs through Supabase.

SCAR's VEX U framework: typed units and geometry, a sensor layer, localization, a command scheduler, and a drivetrain API on top of EZ-Template. 84 tests pass on a PC; it hasn't driven a robot yet.

Task tracking that lives where the team already talks: a Discord bot for /task commands and reminders, and a Next.js board, both on one Postgres database.

Driver control for SCAR's X-drive robot: holonomic mixing of forward, strafe, and turn inputs across four motor groups, with deadbands and power normalization so diagonal moves don't saturate.

A gesture-controlled AR neuroanatomy viewer, built in 24 hours with a team that included Vishal Nagamalla. MediaPipe and OpenCV hand tracking drive a 3D brain model in Unity.