I am building efficient machine learning systems at Precision Numerics.
From 2025 to 2026, I was a research engineer at Samsung Research. I worked on low-precision LLM training and inference: sub-1-bit LLM post-training quantization (NanoQuant), 2-bit LLM quantization-aware training (RaBiT), and GPU kernels for sub-1-bit LLMs (LittleBit).
I was born in California, raised in New York City and Seoul, and currently live in Seoul.
Precision Numerics,
blog demo code 🤗model 💬chat in browser
Boole-1.7B is 13.9x smaller, 1-bit Qwen3-1.7B model. It overall outperforms PrismML's 1-bit Bonsai-1.7B and and state-of-the-art small language models on 10 different math, code, reasoning, and tool-calling tasks.
International Conference on Machine Learning (ICML),
NanoQuant uses low-rank binary factorization for sub-1-bit LLM post-training quantization. It compresses Llama 2 70B by 25.8× in 13 hours on one H100, enabling inference on a consumer 8 GB GPU.
International Conference on Machine Learning (ICML),
RaBiT uses quantization-aware training (QAT), residual-aware optimization, and a dual-binary scheme to produce accurate, efficient binary large language models with performance comparable with state-of-the-art 2-bit quantization methods.