About
I'm Minseo Choi, studying Computer Science at Johns Hopkins University (BS, expected May 2028).
I'm interested in ML systems — the intersection of machine learning, compilers, and GPU performance engineering. Rather than focusing only on model accuracy, I care about how models run in practice: how inference can be faster, cheaper, and more scalable through better systems design.
experience
Research Assistant · JHU Data Science & AI Institute — BLAB
Profiled decode-bound rewriting workloads in a large-scale data-generation pipeline, found high copy-fraction, and deployed draft-free suffix decoding — 1.43× throughput with no extra training or VRAM.
suffix decoding · speculative decoding · GPU profiling
Research Assistant · Johns Hopkins Medicine — PALS Lab
70B generator + five critic agents served with TensorRT-LLM and Triton Inference Server. Fixed orchestration-bound GPU underutilization via generator isolation, async BLS, and KV-cache block reuse — ~3,100 conversations/hour at ~3.3s P95. Now extending to heterogeneous SGLang + TensorRT-LLM serving. With NVIDIA Safety & JHTV; ICML 2027 submission planned.
TensorRT-LLM · Triton Inference Server · SGLang · FP8 · B200
Course Assistant · JHU — Computer Systems Fundamentals
Cache behavior, assembly, and performance debugging in C/C++.
Drill Instructor & Senior Squad Leader · Republic of Korea Army
Trained and mentored 2,000+ recruits.
also building
- GPU kernels — FlashAttention-2 from scratch in Triton/CUDA (~2× speedup, ~5× less peak memory vs naive attention), and fused Triton kernels for medical imaging (~9.7× over unfused PyTorch ops).
- ML compilers — a toy LLVM compiler with IR/CFG visualization, and an end-to-end MLIR pipeline: parsing, passes, toolchain integration.
- This blog — 119 posts on ML systems, GPU runtimes, and AI compiler infrastructure, written since 2024.
long-term
I want to work on inference engines, GPU-accelerated systems, and compiler-driven optimizations that make modern ML models practical at scale. I'm early in this journey — which means there's always more to learn, break, and fix.