Alan Liu

About Alan Liu

I’m a Machine Learning Engineer focused on large-scale language models, distributed training systems, and reinforcement learning for alignment, currently working at Amazon in Seattle.My work centers on building and scaling foundation models across both pre-training and post-training (RL). I’ve led and contributed to end-to-end LLM systems, from model architecture and MoE optimization to large-scale GPU training and kernel-level performance tuning. This includes distributed training on clusters with thousands of GPUs, improving MFU through straggler mitigation and communication–computation overlap, integrating FP8 training, and delivering significant gains in training efficiency, memory usage, and inference speed.On the post-training side, I work on asynchronous reinforcement learning frameworks for LLM alignment, addressing real-world challenges such as training–inference mismatch, policy staleness, and large-scale rollout collection. I have hands-on experience with PPO-style algorithms, GRPO, and off-policy RL at production scale.Before Amazon, I was a Lead Software Engineer at Aptiv, where I designed and optimized high-performance, low-latency LSTM-based time-series models for automotive systems, achieving strong predictive accuracy while reducing inference latency from microseconds to nanoseconds through hardware acceleration and system-level optimizations.I hold a Ph.D. in Computer Engineering from the University of Michigan and have published 20+ peer-reviewed papers with 800+ citations, including first-author work at MLSys. My interests lie at the intersection of LLMs, systems, performance optimization, and scalable ML infrastructure.

Experience

1. Machine Learning Engineer

   Amazon

   May 2024 — Present · Seattle, WA, US

   My work centers on building and scaling foundation models across both pre-training and post-training (RL). I’ve led and contributed to end-to-end LLM systems, from model architecture and MoE optimization to large-scale GPU training and kernel-level performance tuning. This includes distributed training on clusters with thousands of GPUs, improving MFU through straggler mitigation and communication–computation overlap, integrating FP8 training, and delivering significant gains in training efficiency, memory usage, and inference speed.On the post-training side, I work on asynchronous reinforcement learning frameworks for LLM alignment, addressing real-world challenges such as training–inference mismatch, policy staleness, and large-scale rollout collection. I have hands-on experience with PPO-style algorithms, GRPO, and off-policy RL at production scale.

Education

• Beihang University

  Bachelor's degree, Electrical and Computer Engineering

  2012 — 2016

• University of Michigan

  Ph.D, Computer engineer

  2017 — 2022