I don't paste model output into a repo. I turn it into a pipeline that can be measured, rerun, and verified at runtime — then I ship the consumer apps those pipelines feed.
AI TOP 100 (CAMPUS) — Finalist · Apr 2026
Kakao Impact × Brian Impact. Ministry of Science and ICT & Kakao.
Selected as one of 100 finalists from 3,000 students. 본선 진출 · 100인 선정.
Oracle Tarot Paid LLM service. Deterministic engine, multi-provider fallback, billing recovery. |
Haebari Astrology app. Live on Toss, App Store under review. |
Kotoba Offline Japanese. ONNX neural TTS + FSRS v5. |
RelayCode Remote-control local Codex from a phone. Repos and credentials stay on the Mac. |
Vesper — state-based sprites, deterministic assembly, Godot runtime QA
| Project | What it is |
|---|---|
| PixelForge | AI illustration → in-game pixel sprites. UE5 bridge |
| AssetForge | Deterministic 2D asset normalization, validation, and engine export. |
| CutAI | Local AI video editor driven by natural-language cuts. |
| Fissh | Claude Code from a phone. QR pairing, Tailscale auth. |
Build notes live on minseo.log.
problem → measurable baseline → deterministic pipeline → runtime QA → production feedback