AI-powered retrieval layer for everything you save on the web. Designed a global content cache so the full AI pipeline runs once per URL regardless of user count — dropping per-save cost from $0.0024 to ~$0 on cache hit. Exploring dual embeddings for hybrid semantic search and LTREE for AI-generated category hierarchies.

Learned how to keep a map-heavy UI smooth while GraphQL refetched under the hood — built a custom hook that conditionally returns cached vs. fresh data to eliminate flicker during zoom. First time I really felt the weight of making the right abstraction early.

Hobby project. My first time wiring up Socket.IO and a Minimax AI opponent together. Learned how much clean separation between game state and transport layer matters once real-time events enter the picture.

Hobby project. Pure Canvas API and math — a proximity graph where particles form user-typed text. Good reminder that great interactions don't need a framework.