I build small AI systems that run like products: scheduled agents, static publishing, cost-aware inference, and practical data pipelines.
A public lab. The projects below are the actual demos — they run on the same architecture pattern they're described with.
Small AI systems I'm running in the open. One is live with a daily-updating demo; one is in active research. Each has its own page in the Lab — these cards are a quick orientation, not a portfolio.
Patch notes for humans — a daily AI ecosystem change tracker.
A scheduled agent collects, filters, summarizes, translates, and scores AI/developer updates each morning, then publishes static JSON this page renders. Keyword filters, EN/KO, and a usefulness score per item — and no LLM on your request path.
An AI-powered pole dance learning & choreography platform built on a skill knowledge graph.
Models pole dance moves, prerequisites, transitions, difficulty levels, and learning paths as a knowledge graph. Users discover optimal progression routes toward a target skill, generate personalized combos from their current abilities, and explore move relationships through an interactive visual interface.
Current status: Research, architecture design, graph modeling, and prototype planning.
The same architecture shows up across every demo here. It is the actual reason the site is hostable from a tiny VPS and survives a flaky upstream.
A cron agent runs the LLM work upstream — never on your request. The site is static.
The agent writes a single JSON artifact. The site reads that file. That's the whole contract.
A bad upstream or a flaky LLM only delays the next refresh. It can't break the page in front of you.
Filter and dedup before the LLM. Route easy items to a cheap model. Cost stays small and visible.
For the projects above, or collaboration on similar systems.