Harim Choi
Machine Learning Engineer · Seoul, South Korea
I am a machine learning engineer who quickly understands new problems and solves them through model research and systems development. Across 7+ years, I have worked on data analysis, workflow automation, and prediction systems across public procurement, construction, and retail. I have improved model training methods and GPU kernels, taking responsibility for data preparation, model deployment, and operations.
Highlights
- State-of-the-art weakly supervised semantic segmentationCOCO-Val mIoU 53.31%
- INT2·INT3 PTQ SOTA in LLM quantizationLlama-3 8B INT3 WikiText-2 6.99 · 7.36× decoding
- MCP search server with automatic knowledge graph constructionTop 1% on MCP TOPLIST · 291 GitHub stars
Recent writing
- WarpQuant: Dual-Domain Low-Bit LLM Quantization
A post-training quantization method that reaches INT2/INT3 SOTA accuracy and runs the compressed weights directly through a C++/CUDA llama.cpp kernel — 7.36× faster decoding than GPTQ W3.
- When to Use What: The Missing Framework for Multi-Agent Competitive Systems
Four questions for choosing a method in competitive multi-agent problems such as sealed-bid auctions, ad bidding, and trading.
- Building a Production LangGraph Travel Agent
How I built a LangGraph travel assistant that searches flights, hotels, and activities in parallel and assembles budget-aware packages.
Experience
- Sejong Bid InstituteResearcherApr 2026–Aug 2026
- Hanmac GroupML Engineer / ResearcherNov 2024–Feb 2026
- Sanha General TechnologyAI / Automation EngineerJan 2021–Nov 2024
- Nielsen Korea GTCData AnalystAug 2019–Oct 2020