Projects
Two decades of work at the boundary of physics, biology and machine learning — from simulating blood flow in the human heart and mapping memory circuits in the brain, to computer vision and continual learning at Facebook AI, to the long-horizon coding agents I build at Google today.
The common thread is building systems that learn from messy, real-world signal and then survive contact with production: an FDA-cleared surgical planner, alt text for every photo on Facebook, a continual learning platform serving Instagram ranking and integrity, and agents that write and tune TPU kernels. The five areas below collect the projects, talks and publications for each, newest first.
Research Areas
Coding and Reasoning LLMs
AI-4-AI: post-training, reinforcement learning and agentic workflows for coding agents that discover and maximize the performance of AI algorithms on accelerators. MaxKernel (NeurIPS 2026 spotlight), JAXBench, MLE Agent, Gemini Code Assist and GPU-cluster degradation forecasting.
Continual Learning
Tech lead for Continual Learning at Scale. CLAP, a PyTorch platform bringing continuous sampling, prompt-based learning and memory rehearsal to production models — 3–10x recall gains across integrity, ranking and misinformation — plus Refiner Fusion Networks for multimodal self-supervision.
Computer Vision
AI tech lead for Automatic Alt Text, which automatically describes photos for blind and visually impaired users of Facebook. Semi-supervised cluster-based and active learning that discovered thousands of visual concepts at scale.
Computational Neuroscience
How are memories stored and retrieved? Computational tools for in-vivo two-photon imaging of behaving mice showed that memory is encoded as a small-world network, revealing hub neurons preferentially recruited during retrieval. Published in Nature as co-first author.
BioPhysics and AI
Fifteen years combining physics-based simulation with machine learning: real-time patient-specific blood flow, an FDA-cleared surgical stent planner, AI for cardiovascular imaging (HALE stenosis quantification, automated image quality), and uncertainty quantification for stochastic PDEs.
Selected Outcomes
Nature, co-first author
Top-down control of memory retrieval and the discovery of hub neurons, with the Deisseroth Lab at Stanford.
FDA-cleared in the clinic
A real-time stent planner used by physicians in the United States and abroad, plus stenosis quantification used routinely to diagnose coronary artery disease.
Deployed at billion-user scale
Automatic Alt Text for blind and visually impaired users, and CLAP continual learning across Facebook and Instagram.
By Period
| 2023–present | Google — Coding and Reasoning LLMs |
| 2020–2023 | Facebook AI — Continual Learning, Computer Vision |
| 2012–2019 | HeartFlow — BioPhysics and AI |
| 2013–2014 | Stanford University — Computational Neuroscience |
| 2004–2012 | UC San Diego · Cornell University — BioPhysics and AI |