Personal and research projects. Most are public on GitHub. Documentation detailing the process is written if you want to learn more.
★ 1st Place Grand Prize · UC Berkeley AI Hackathon
Turns physical spaces into navigable 3D scenes you can explore and query in natural language. A COLMAP + Gaussian Splatting + WebGL pipeline reconstructs a walkable scene from photos in ~4 minutes, paired with a spatial AI agent that drives the camera to answer questions about the space.
A computer vision pipeline that reconstructs 2D Clash Royale screen recordings into animated 3D arena replays. Fine-tuned YOLO + ByteTrack to track each troop's position and identity, then rendered keyframed 3D animations in Blender via Python.
A multi-agent AI security assessment tool that runs staged adversarial attacks against AI products (recon, prompt extraction, data exfiltration, privilege escalation), streaming live agent activity to a dashboard with an executive report.
A campus events discovery app unifying scraped event data from 5 web sources into Supabase, powering a React Native feed with search, filters, and a campus map for browsing upcoming events.
Supervised models trained on engineered match-level features to predict Premier League outcomes across a full 380-game season, hitting 61.5% accuracy vs a 33% baseline. Published in NHSJS.