Haibin Lin is a research scientist and engineering leader with 12 years of experience building foundation-model infrastructure and large-scale ML systems from the Bay Area. He leads a ~30-person team at ByteDance Seed focused on training infrastructure for search, recommendation, moderation, multi-modal models and LLMs, and has co-authored papers at top systems and ML venues. Previously at AWS he helped optimize MXNet and distributed training stacks, delivered the fastest BERT pre-training in the cloud at the time, and contributed to production NLP toolkits used across Alexa and other services. A prolific open-source contributor, his work spans deep learning frameworks and compilers (MXNet, TVM, DGL, GluonNLP, mshadow) with practical low-level optimizations in sparse ops, memory planning and distributed training. He blends research and engineering: comfortable in compiler-level C++/CUDA fixes as well as RL/LLM training pipelines and MLOps integrations. An uncommon strength is his track record of moving cutting-edge research into robust, production-grade infrastructure that scales.
12 years of coding experience
4 years of employment as a software developer
The University of Hong Kong (HKU)
Exchange Program Computer Science, Exchange Program Computer Science at University of Toronto
Master’s Degree Computer Science, Master’s Degree Computer Science at Carnegie Mellon University
Electrical and Electronics Engineering, Electrical and Electronics Engineering at Shanghai Jiao Tong University
Contributions:11 releases, 1 review, 151 commits in 2 years 2 months
Contributions summary:Haibin made significant contributions to the `gluon-nlp` repository, demonstrating a strong focus on Natural Language Processing (NLP) tasks. Their work included enhancements to the data stream API, specifically for streaming corpora, by adding support for custom samplers. Furthermore, they were deeply involved in the development of BERT models, including bug fixes, updates, and optimizations. Finally, the user added a pre-training script for BERT using the OpenWebText dataset.
verl/HybridFlow: A Flexible and Efficient RL Post-Training Framework
Role in this project:
ML Engineer
Contributions:5 releases, 545 reviews, 411 PRs in 9 months
Contributions summary:Haibin primarily contributed to the Verl (Volcano Engine Reinforcement Learning for LLMs) project, focusing on enhancements related to reinforcement learning, especially in the context of LLMs. Their work included making the `wandb` dependency optional, which improves modularity. They modified various scripts associated with PPO (Proximal Policy Optimization) training, notably improving flexibility by allowing additional arguments to be passed. These changes suggest a focus on refining the training environment and the tools used for RL experiments.
post-training
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.