Location: San Francisco or Croatia
Type: Full-time
Reports to: CTO
The role
You will lead the pod building Daytona's serverless GPU sandboxes: full-passthrough GPU compute with the same instant-start economics that define Daytona everywhere else. Pause, snapshot, and fork for GPU-attached workloads is a hard, partially solved problem - we have a working checkpoint/restore path, and you will take it to production. Customers use these sandboxes for agent-driven ML workloads, RL training, inference experimentation, and anything that needs a real GPU behind an API call.
This is an engineering role first. You will spend most of your week shipping code on the critical path and owning the pod’s technical direction, and you will also manage the pod: 4 senior engineers working alongside a dedicated PM. Management is part of the job, not the whole job.
What You'll Do
- Own architecture and delivery for the GPU sandbox runtime as we pivot from MIG-based sandboxes on our H100 fleet to full VFIO passthrough: KVM/QEMU device passthrough, IOMMU isolation, driver lifecycle, and guest image management
- Drive fleet efficiency: bin-packing, cold-start latency, GPU utilization, and the economics of serverless GPU (idle cost is the enemy)
- Own reliability and security posture for multi-tenant GPU hosts, including isolation guarantees customers can put in a security review
- Work with our bare metal and cloud providers on H100-class capacity, hardware quals, and fleet expansion
- Hire, grow, and unblock the pod; set the bar for engineering quality
- Partner with the pod PM on roadmap, pricing inputs, and customer-facing commitments (SLAs, quotas, regions)
What We're Looking For
- 8+ years in systems or infrastructure engineering, 2+ years leading engineers who ship
- Deep hands-on experience with Linux virtualization: KVM, QEMU, libvirt, VFIO/PCIe passthrough, IOMMU
- Production experience running GPU infrastructure at fleet scale (scheduling, health, driver hell, RMA workflows)
- Strong systems programming background (Go, Rust, or C/C++)