Heungsub Lee
Server Developer at Toss
Yongin-si, South Korea
Summary
🤩
Rockstar🎓
Top SchoolHeungsub Lee is a server developer with 13 years of experience building high-throughput distributed systems, from MMORPG backends that handled tens of thousands of concurrent users to large-scale ML infrastructure for Korean LLMs. He has led engineering teams (up to 25 engineers) and shipped platform capabilities like NSML v2 and MLOps optimizations for HyperCLOVA, blending hands-on coding with people leadership. His open-source contributions span Python ecosystem projects—fixing Babel extraction edge cases, hardening pyzmq tests, and implementing a TrueSkill library—plus a notable torchgpipe implementation for pipeline-parallel training in PyTorch. Practically minded, he pairs rigorous test automation and debugging skills with production engineering: improving logging, adding test coverage, and isolating hard-to-reproduce concurrency bugs. Based in Yongin-si, South Korea, he moves fluidly between backend systems, ML tooling, and performance engineering, with a track record of making complex, large-scale systems more reliable and observable.
13 years of coding experience
18 years of employment as a software developer
Software Engineering, Software Engineering at Kwangwoon UniversityDesign Programming, Design Programming at Sunrin Internet High School
English, Japanese, Korean
Stackoverflow
188reputation
15kreached
6answers
1question
Github Skills (33)
debugging10
pytorch10
debug10
zeromq10
checkpoint10
rails-i18n10
pytest10
parallelization10
python10
multithreading10
testing10
machine-learning10
checkpointing10
gevent10
deep-learning10
Programming languages (15)
Github contributions (5)
kakaobrain/torchgpipe
May 2019 - Sep 2020
A GPipe implementation in PyTorch
Role in this project:
Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:9 releases, 3 reviews, 330 commits in 1 year 4 months
Contributions summary:Heungsub primarily focused on implementing the GPipe architecture in PyTorch for deep learning model parallelization. They developed the core GPipe functionality, including micro-batching, pipeline execution, and checkpointing mechanisms. The user also integrated support for skip connections and handled the complexities of managing random number generator states within the checkpointing process. Moreover, the user contributed to making the module usable by ensuring the module handles for various corner cases such as none skip.
pytorchdeep-learningtorchdistributed-traininggpipe
what-studio/profiling
Sep 2014 - Aug 2020
An interactive continuous Python profiler.
Role in this project:Back-end Developer
Back-end DeveloperContributions:2 releases, 422 commits, 15 PRs in 6 years
Contributions summary:Heungsub contributed to the development of an interactive Python profiler. Their work included implementing core features such as the ability to keep focus when sorting keys, specifying the version of a dependency (urwid-geventloop), and correcting a bug related to frame recording. They were also responsible for adding threading support to the profiler.
pythoncontinuouscython
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.