Won 1st Place in the Nano Banana track at Google DeepMind x UCLA. Type a scene, then edit and animate the result in the browser. Gemini 2.5 plans the layout, NVIDIA Kimodo creates the character motion, and Nanobanana with Trellis turns images into textured 3D models. The editor includes a motion timeline, transitions, and video export.

Won 1st Place at the IBM SkillsBuild x OSU AI Hackathon. Sparky helps students build and debug breadboard circuits. Its agent can place components, inspect its work, fix wiring mistakes, and explain each change. I built the circuit simulator in Three.js and added accounts so students can save and share their work.

Describe a part in plain English and get an editable 3D model with controls for each dimension. A local Claude Code session writes OpenSCAD, then OpenSCAD WASM compiles and renders it in the browser. Models can be exported as STL, SCAD, or DXF. There is also a benchmark for comparing model output on the same CAD prompt.

A donor research tool for nonprofits, built in one day at the Datadog Agentic Engineering Hackathon. Search for a cause in plain English and Tribe checks 2.8 million public FEC records for people who have supported similar work. Results are ranked, enriched with current web data, and linked to the original filings so the reasoning is easy to verify.

Won 1st Place Overall at the BDAA x Lovable Hackathon. A campus lost and found that verifies ownership before returning an item. Gemini creates questions from the original photo that only the owner should be able to answer. A Mapbox view shows where items were found, and fuzzy search helps students find close matches.

A Chrome extension I use for my own laptop resale business. It collects Amazon listing data across multiple pages and exports it to Excel, replacing hours of manual copying. More than 200 users have installed it, and Google gave it the Featured badge on the Chrome Web Store.

I built this to check whether furniture would fit in a New York apartment from a video tour. It turns a phone walkthrough into a measurable 3D point cloud. VGGT reconstructs the room, while UniDepth and known object sizes correct the scale. You can measure walls and place furniture models directly in the result.

I made Melody so I could listen to music on the subway without a connection or another subscription. Songs are stored in IndexedDB and play fully offline. It supports playlists, favorites, fuzzy search, synced lyrics, cover art, lock screen controls, and home screen installation on iPhone.

I compared 2.15 million NYC collision records with more than 3,300 bus stops to identify the riskiest places to wait. GeoPandas linked each crash to its nearest stop and ranked the ten highest risk locations. The project was selected from 400+ applicants for presentation to the Department of Transportation and Federal Highway Administration.

Point a camera at a document, sign, or screen and ask questions about what it says. Google Cloud Vision reads the text and Perplexity answers with relevant context. The split screen keeps the captured text beside the answer, so it is easy to check the result.