Youngmin Ko · AI Forward Deployed Engineer at KRAFTON.
Research, open source, and the evidence behind the work.
At KRAFTON I combine domain rules and AI for business workflows, connecting approved inputs to execution outcomes. In research I check whether evaluation numbers beat chance.
| Employment | KRAFTON · AI Forward Deployed EngineerPDF |
|---|---|
| Paper | Disagreement between time-series anomaly-detection metrics · work in progress↗ |
| Open source | |
| Degree | B.S. in Information Sciences and Technology, Penn State · GPA 3.68 / 3.87 major · Dean's List(not public) |
| Based in | Seoul, South Korea |
Every number here links to its source. Where no public artefact exists, the claim is marked (not public).
NeurIPS 2026 Workshop
Preprint
The evaluation toolkit behind the paper's experiments. Null models separate real rank flips from chance, and a cluster-aware bootstrap puts confidence intervals on the rank-flip statistics.
Developed and operated meeting workflows connecting recording, transcription, source verification, reviewed minutes and follow-up actions. Built internal customer-support workflows connecting privacy masking, official knowledge, response review and agent handoff.
Operators judged link trouble between unmanned ground vehicles and their control station by feel. Alone in eight weeks, built a detector that turns that call into latency, jitter and packet-loss signals. With no labeled faults, rules decide from day one and an Isolation Forest joins as data accumulates; every alert shows which signal moved and by how much, so an operator can check it.
Re-architected website structure using code minification, lazy loading, and image compression to improve page load performance
Open PR for review, not merged: opt-in cleanup for delivered Python processes in the existing FDE Gym runner. Requires macOS 15.4+ with original-parent generation support verified at runtime; unsupported hosts are rejected before delivered code runs. In the bounded detached-child comparison, all 184 per-case results stayed unchanged while observed requests reaching later incidents fell from 60 to 0. This is lifecycle cleanup, not a complete containment boundary; no adoption or score gain is claimed.
Preserves original intent and compares patching, reframing, and rebuilding against the same browser checks. Feature expansion is on hold while the public pilot tests actual use; external adoption is unverified.
About a hundred servers claimed to be Jev-compatible with nothing to check the claim against. I turned scattered official docs into a 48-requirement spec with tests and ran it on the eight most-starred servers: two passed. Maintainers of four shipped fixes within three days of the reports.
Education
GPA 3.68 / Major 3.87 · Dean's List(not public)
Hackathons & Competitions
EquiLaw: 20th / 1,500 teams↗
Top 4.2% / 1,500 teams(not public)
Top 5.5% / 1,400 teams↗
Activities & courses
Invited to the FDE practitioners' panel at KakaoBank's KodeRunner 2026, alongside FDEs from OpenAI and Toss. The session was on solving business problems with AI and making the results reliable.