I keep large training runs alive.
For the last seven years that has meant pathology foundation models at Paige.AI and now Tempus. Before that, it meant the Compact Muon Solenoid experiment at CERN’s Large Hadron Collider, where I ran the Caltech Tier-2 — one of the U.S. compute sites the physics collaboration depends on to reconstruct collision events. Before that, enterprise systems at Duke and Unix production at FedEx.
The thread is not “AI” or “physics.” It is production infrastructure for missions where the cost of a failed run is high — months of physicist time, or weeks of GPU-hours on a billion-parameter model. The skills transfer better than people assume. Scheduling, storage tiering, network fabric, observability, change management under pressure — the LHC people and the foundation-model people are solving the same problem at different scales of data and different shapes of compute.
What I’m working on now is the migration of Paige’s pathology training and inference into Tempus’s GCP environment, while continuing to scale the existing Azure footprint. The interesting problems are around checkpoint resharding across heterogeneous fabrics, training reliability at multi-node scale, and storage architectures that can hold 7 PB of whole-slide imaging hot enough for randomized I/O during training.
Outside Tempus I am most interested in the infrastructure layer of biological and scientific foundation models — autonomous wet-lab platforms, structural biology training pipelines, and the next generation of clinical AI where the data is larger than the model. If that is what you’re building, say hello.
The CV has the full chronology. Publications and acknowledgements here.