Sethu Sankaran

I am an AI researcher and engineering leader at Google, where I lead work on long-horizon coding and reasoning tasks. Over 20 years in industry and academia, my work has combined learned models with the structure of the problem, whether that is code, images, streaming data, physics or biology, to build AI systems that work at scale:

• Coding and reasoning agents. At Google: post-training, reinforcement learning and long-horizon coding agents that discover and optimize AI kernels, including JAXBench (ICML DL4Code), MaxKernel (NeurIPS 2026 spotlight), MLE Agent and Gemini Code Assist. This work is a part of kernel coding agents in Google Antigravity.
• Computer vision. At Facebook AI: AI tech lead for Automatic Alt Text, which describes photos for blind and visually impaired users, using semi-supervised and active learning to discover thousands of visual concepts.
• Continual learning. At Facebook AI: CLAP, a platform that keeps production models learning from fresh data, with 3–10x recall gains across Facebook and Instagram balancing forgetting and intransigence.
• Computational neuroscience. At Stanford: graph analysis of in-vivo brain imaging that revealed hub neurons in memory retrieval, published in Nature as co-first author.
• Physics-informed AI. At HeartFlow: physics-driven and machine-learned models of blood flow in the heart, including an FDA-cleared real-time surgical planner and AI for cardiac imaging.

I received my BS from IIT Madras, my PhD from Cornell University and did postdoctoral research at UC San Diego and Stanford University.