| Location | San Francisco, CA |
Location: San Francisco
Type: Full-time
Reports to: Engineering Lead, GPU Sandboxes
The role
You will build the runtime that gives every AI agent its own GPU. Full passthrough, hard isolation, serverless economics. This is deep systems work: hypervisors, kernels, drivers, and the orchestration layer that makes an GPUs appear behind an API call in seconds.
What You\'ll Do
Build and harden the GPU sandbox runtime: KVM/QEMU, VFIO passthrough, IOMMU group management, guest driver and CUDA stack lifecycle
Attack cold-start latency across the stack: image caching, VM boot paths, driver init, snapshot restore
Build fleet-level systems: GPU-aware scheduling and bin-packing, health monitoring, failure detection, automated remediation
Implement pause, snapshot, and fork semantics for GPU-attached workloads
Own security-relevant surfaces: device isolation, side-channel considerations, tenant boundaries on shared hosts
Debug the ugly stuff: PCIe errors, IOMMU faults, driver crashes, thermal throttling, the works
What We\'re Looking For
5+ years of systems engineering in production environments
Hands-on depth in Linux virtualization internals: KVM, QEMU, libvirt, VFIO, IOMMU
Strong Go, Rust, or C/C++ and real Linux kernel-adjacent debugging skills
Experience operating GPUs in production (NVIDIA data center hardware preferred)
You ship, you own your services, and you like hard problems more than big teams
Nice to have
Contributions to QEMU, KVM, Firecracker, Cloud Hypervisor, or similar
NVML/DCGM tooling, GPU telemetry, or driver-level work