Forward Deployed Engineer at KRAFTON, Seoul. I put AI to work inside real teams and measure it before anyone relies on it.
Highlights
- Took a marketing-reporting tool into internal production: data validation and server stabilization; the team estimated its monthly reporting work fell from 60 hours to 5
- Lead developer of an internal meeting-automation service: recording, transcription and minutes, from hackathon prototype to internal launch
- Data and evaluation for a player-support AI: labeled 31,000+ Chinese tickets (87% in a random audit) and analyzed 1.44M tickets to size what AI should take on
- Sole-author research: the widely cited 31% metric disagreement is below its 50% chance level · project page
- Invited panelist, KodeRunner 2026 by KakaoBank: an FDE panel with OpenAI and Toss · KakaoBank tech blog
Open source
| jobradar | Reads job postings in full and grades them against your own rubric with Claude Code |
| sessionreel | Claude Code session log → 30–60 second recap video · PyPI |
| claude-drift | Replays your Claude Code sessions on a new model to show what changed · PyPI |
| jevcompat | Conformance suite for Jev-compatible API servers · four open-source servers shipped fixes from its reports within three days |



