Projects grouped by what they're about — badges show maturity (Live · Beta · Concept).
WorshipFlow — run a worship service end to end: an 867-song library with versions, song forms and YouTube references, setlist (콘티) building, and a score viewer that follows the setlist order on stage.
An honest data-science study of the Korean lottery: a 9-stage curriculum (probability → testing → Bayesian → ML → unsupervised → backtest) whose conclusion is that no usable signal exists. Not a prediction tool — a proof of why prediction fails.
A local-first AI orchestrator built to get real work out of cheap open-weight models. Its generalist tiers resolve to Llama 3.1 8B and 70B, so the engineering budget goes into keyword-first routing that spends almost no tokens, typed task decomposition with a halt-on-unverified guardrail, lexical-first retrieval that carries citations, and a UI that makes the cost of every decision visible.
A closed-loop optimizer over housing. Knowledge-acquisition and design-mutation actions compete in one priority queue under one currency — expected decision-improvement per cost. The loop is the product; solvers, adapters and UI are plug-ins.
A knowledge base for running a Korean social cooperative — official-source tracking with change detection, lifecycle checklists, hybrid keyword + vector search, and AI drafts restricted to approved evidence (opt-in per document).