Portrait of William Hu

William Hu (Haoqi Hu)

Senior Applied Scientist at Amazon

Bellevue, WA

About

I'm a Senior Applied Scientist at Amazon on the POE-AI team, where I work on large language models for seller-facing applications. That means carrying a model through its whole lifecycle, from continued pretraining and instruction fine-tuning through reinforcement learning and evaluation, and caring about the data recipes feeding it as much as the systems that eventually serve it.

Much of that comes down to distributed training. I've built and run pipelines across most of the common ways of splitting a job over many machines, whether that means plain data parallelism, sharding optimizer state and parameters across ranks, or carving a model up along tensor and pipeline boundaries. In practice that has meant PyTorch DDP and FSDP, DeepSpeed, and Megatron-LM, running on SageMaker HyperPod clusters under Slurm and EKS.

My research circles the same problems from a different angle: what to train on, how to train it, and how to tell whether it worked. That pulls me toward data selection, training methods, evaluation and benchmark design, robustness, and increasingly multimodal models. I publish occasionally, sometimes with academic collaborators. Before Amazon I spent a few years as a deep learning engineer at the Bosch Center for Artificial Intelligence, after an M.S. at Carnegie Mellon and a B.S. at Ohio State.

Away from all of this I follow markets and investing, and I enjoy thinking through startup ideas. I'm a blue belt in Brazilian jiu-jitsu and play more tennis than I probably should. Happy to talk about any of it, so do get in touch.

Publications

Academic Service

Reviewer for IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) in 2026, PLOS ONE in 2024, and the Multimodal Algorithmic Reasoning (MAR) Workshop at CVPR 2024.

Experience

Education