Forward Deployed Engineer at KRAFTON, Seoul. Before that, research engineer intern at Hanwha Aerospace; B.S. from Penn State.
Work
- Meeting automation. I lead development and maintenance of a product connecting recording and transcription, source playback and speaker verification, personalized minutes, approved sharing, and owner-linked follow-up actions. Persistent approval and execution state support resuming the same meeting workflow.
- Customer support. I implemented a service connecting privacy masking and ticket classification, official-knowledge retrieval and response drafts, agent review and handoff, and case-level outcome tracking.
- Advertising reporting. I contribute data validation, authentication, deployment migration, and operational stabilization to a business-owned reporting product. My work improves aggregation, caching, title registration, and visibility into execution and delivery.
- Invited panelist at KodeRunner 2026 (KakaoBank), alongside FDEs from OpenAI and Toss · KakaoBank tech blog
Research
- What Ten Seeds Can and Cannot Resolve About Replay Strategy Selection Under Distribution Shift, accepted at a NeurIPS 2026 workshop (CL4FMAgents). Sole author · code
- The 31% rank disagreement between time-series anomaly detection metrics is often cited as a reason to change how detectors are scored. Two unrelated metrics disagree 50% of the time by chance, and where both metrics clearly separate two detectors the rate is 2%. Sole author · project page
Open source
| jevcompat | A 48-requirement spec and conformance tests for Jev-compatible API servers. Two of the eight most-used servers passed; four maintainers shipped fixes within three days of its reports. |
| sessionreel | Turns a coding-agent session log into a one-minute recap and flags runs that only looked successful · PyPI |