Pratyay Pandey

Hi, I'm Pratyay. I work on the structure of intelligence: how concepts are organized inside models, and the systems that make models fast.

I finished my Berkeley EECS degree in a year and a half, mostly so I could spend all my time on research. Currently that means two things. At BAIR, I work on mechanistic interpretability under Trevor Darrell and Jacob Steinhardt, studying steering vectors, training data attribution, and the geometry of concepts in representation space. At Databricks, I'm one of two non-PhD interns on the LLM inference research team. In my first three weeks there I enabled hybrid speculative decoding in our internal SGLang, one of the largest architectural changes since the framework's creation, and I was part of the small team that served GLM 5.2 at 381 tok/s within 24 hours of its release, the fastest in industry at the time.

Previously: research at Sky Computing Lab under Matei Zaharia, two semesters TAing CS162 for Ion Stoica and Matei Zaharia, and Eregion, a dynamic fine-tuning pipeline that cut GPU usage roughly 60%, won the 1517 Fund's Medici Grant, and was pitched to Anthropic.

Outside of research, I write poetry, learn Hindi love songs on the piano or guitar, and have strong opinions on coffee. If you think we'd have an interesting conversation, email me.