EN · KO
Summary
M.S. in Artificial Intelligence (SNU) specializing in LLM-based code generation and evaluation; 3 papers at ACL / COLING (second author on ArchCode and DuET, fourth author on PERC; ArchCode 37 citations).
Sole implementer of the experimental frameworks behind two of them, including evaluation harnesses that run generated code in Docker-isolated, identical environments to measure not just correctness but execution time and memory fairly.
Hands-on with LLM serving (vLLM, Ollama, LiteLLM) and agent pipelines (LangChain, LangGraph).
Skills
LLM evaluation harnesses: Evaluation code for 6 code-generation benchmarks; Docker-isolated runs measuring accuracy, execution time, and memory
LLM inference & serving: vLLM, Ollama, LiteLLM
LLM agents: LangChain, LangGraph, RAG, text-to-SQL, multi-agent orchestration
Experiment infrastructure: PyTorch, Docker, multi-GPU Linux server administration
Experience
- Researched LLM-based text-to-SQL and agentic database exploration on medical-domain databases (unpublished): LLM agents explore the schema, actual values, and domain-specific local knowledge over multiple steps to build an answer.
- Also studied how to judge ambiguous cases where an agent’s SQL returns a different result from the gold SQL yet can still be correct.
- Co-developed LLM code-generation and evaluation methods; 3 papers (2nd author: ArchCode, DuET; 4th: PERC); advised by Prof. Seung-won Hwang
- Sole implementer of the ArchCode and DuET experimental frameworks; LLM pipelines built with LangChain/LangGraph
- Built Docker-isolated evaluation harnesses measuring correctness, execution time, and memory fairly; ran all benchmark experiments (HumanEval, CodeContests, MultiPL-E, LiveCodeBench, BigCodeBench, DevEval)
- Researched multi-hop QA and text-to-SQL (SQLBot); TA for a Samsung Big Data course; managed the lab’s multi-GPU servers
- Voice-assistant robot arm; drone-sound landing guidance
- BERT prototype (TensorFlow) for the Bixby capsule classifier
Personal Project
macp (Maestro Agent Control Plane) — autonomous agent operations
- The user just queues tasks; workers analyze and split each one, assign a model tier (heavy / standard / light) and tools by difficulty, and run coding agents through research → plan → implement → verify → report, 24/7. When the task pool runs dry, a planner runs and re-reads the projects to find needed work. In 3 months of real use (2026.7–10), completed 1,080 tasks autonomously across 22 independent project repositories. Orthogonal to any particular coding agent.
Selected Publications
ACL 2024 · 37 citations · SOTA code-generation accuracy; introduced HumanEval-NFR, the first benchmark for non-functional requirements in generated code
Findings of ACL 2026 · SOTA test-output prediction by combining LLM-simulated pseudocode execution with real code execution
COLING 2025 (short) · retrieves few-shot examples by pseudocode plans for low-resource languages; beats prior RAG methods
Education
- Advised by Prof. Seung-won Hwang
- cum laude
- Exchange semesters: UC Berkeley (2017.06–08); Stony Brook University (2017.09–12)