Henry Pan

I work on large-scale LLM post-training, training-system optimizations, and research tooling at Amazon AGI. Previously at Instacart, I led ML projects addressing marketplace problems, from problem framing through model development, experimentation, and production serving. Earlier, I built distributed systems and trading infrastructure.

Independently, I pursue research in AI, including agent harnesses (code, experimental results) that improve LLM capabilities without updating model weights. I like learning from first principles, which led me to build a deep learning library from scratch (code, write-up). I share my research, experiments, and what I learn on my blog.

In a previous life, I was a business major and worked in finance and consulting, but I wanted to build rather than only analyze, hence the switch. More about my background.