Member of Technical Staff, Sandbox Infrastructure

Our GPU compiler searches a very wide space of kernel optimizations and verifies the result at the end. That search only works if we can run large volumes of untrusted, newly generated kernels on real hardware, quickly, and without putting anything else at risk. You will own that layer: a serverless GPU container service spanning NVIDIA, AMD, TPU and Trainium. Its measurements feed the compiler's reward signal, so the sandbox has to be accurate as well as safe. It also carries post-training runs and customer workloads that need hard isolation. What you'll own Extend the sandboxing stack to new vendors and accelerators. Build and maintain GPU virtualization, including gVisor work at the driver and ioctl level. Make sandboxes survive spot capacity through preemption-aware scheduling, checkpointing and rescheduling. Support multi-GPU and multi-node sandboxes across NVLink, NVSwitch, InfiniBand and RoCE.Own live migration, including socket-preserving migration. Guarantee measurement and profiling fidelity under isolation. Keep the compiler, post-training and kernel teams from ever being throughput-limited by sandboxes.

01 Intro call 30 minutes with a founder. What you have built, and what you want to build next.02 Systems deep dive 90 minutes on a system you know cold. No puzzles, no whiteboard trivia.03 Work session Half a day with the team on a real sandbox problem. Paid.04 Offer References, then an offer within a week. Relocation support to San Francisco.