Yaroslav Bulatov
Advisor
Yaroslav was part of the original TensorFlow design team at Google Brain, working one-on-one with Jeff Dean and writing the framework's first Python client. At OpenAI he created gradient checkpointing — now a standard technique for training larger networks within fixed memory — and later led a small team that beat Google at the DAWNBench competition for fastest ImageNet training. He has held principal and research roles at Together AI, Meta (Distributed PyTorch), Imbue, and Contextual AI, building large-scale training infrastructure behind state-of-the-art models, and mentored Ian Goodfellow as an intern early in his career. Today he leads the Sutro Group, a research collective focused on energy-efficient deep learning. In his spare time, Yaroslav tinkers with algorithms and numerical mathematics for fun — an obsession that has produced 850+ Mathematica notebooks and a long trail of contributions across math-minded corners of the internet.
Yaroslav was part of the original TensorFlow design team at Google Brain, working one-on-one with Jeff Dean and writing the framework's first Python client. At OpenAI he created gradient checkpointing — now a standard technique for training larger networks within fixed memory — and later led a small team that beat Google at the DAWNBench competition for fastest ImageNet training. He has held principal and research roles at Together AI, Meta (Distributed PyTorch), Imbue, and Contextual AI, building large-scale training infrastructure behind state-of-the-art models, and mentored Ian Goodfellow as an intern early in his career. Today he leads the Sutro Group, a research collective focused on energy-efficient deep learning. In his spare time, Yaroslav tinkers with algorithms and numerical mathematics for fun — an obsession that has produced 850+ Mathematica notebooks and a long trail of contributions across math-minded corners of the internet.