Projects

Predicts seven physical and chemical properties of organic molecules (melting point, boiling point, heat of fusion, heat of vaporization, critical temperature, critical pressure, and flash point) directly from SMILES structure. Combines RDKit 2D descriptors, Mordred 3D conformer descriptors, and Morgan fingerprints, feeding a CatBoost model tuned per-property with Optuna over a 70/15/15 train/val/test split.

About 3,300 of 4,343 organic compounds retained after filtering. Best result: critical temperature, R² 0.939 (n=2,846). Hardest target: melting point, R² 0.674 (n=3,175), a known limitation since 2D/3D descriptors don't fully capture crystal packing.

A tool for Westwood High School students to see how AP exam scores and Austin Community College dual-credit courses translate into transfer credit at target universities. Pick a school, select exams and courses, and get back estimated credit hours and equivalent course codes. Static front end with per-university credit policy data stored as JSON, no backend, no database, covering nine universities including UT Austin, Texas A&M, and Purdue.

A monorepo analyzing the Hypixel SkyBlock Bazaar, a live, player-driven in-game market, as a real market data problem: ingest, compute, expose. A Go calculation engine polls the bazaar API and maintains rolling per-product metrics (price history, order-flow deltas, moving averages); a Flask dashboard prices out crafting chains against live prices to surface profitable trades; a SvelteKit frontend and Discord bot expose the results.

A Kotlin-based Minecraft server engine in the spirit of Hypixel SkyBlock, with custom game mechanics, server-side systems, and infrastructure built from scratch rather than run on stock server software. I contribute as a named developer under the handle dev-gw, working on NPC systems, gameplay features, and infrastructure, running my own service within the project.

Private repository. Deployment forthcoming.

A semantic search engine over the xkcd comic archive. Type a situation or half-formed thought and get back the comic that fits, even without knowing its title or number. A Scrapy spider builds the corpus from explainxkcd.com; queries are expanded by a local LLM (Ollama, llama3.2), then searched via hybrid FAISS + BM25 retrieval fused with Reciprocal Rank Fusion.