Software Engineer, Cloud Infrastructure

About EdgeMatrix EdgeMatrix builds cloud infrastructure for AI workloads, from training clusters down to inference at the edge. We're at the point where the interesting problems are no longer about standing up GPUs, they're about running heterogeneous compute reliably across a lot of tenants with very different needs, and doing it without the control plane becoming a mess. Our engineering culture is small, senior, and writing-heavy. We prefer design docs and thorough review over standups, and we ship in reasonable increments rather than sprint theater. The Role We're looking for a software engineer to own the virtualized compute layer of our platform. This is the system that turns higher-level intent into actual running workloads, and it sits between our customer-facing APIs and the physical fleet. You'll design the lifecycle API that the rest of the platform builds against, and you'll be responsible for making it work cleanly across multiple backends, from full VMs for standard tenants to lightweight microVMs for sandboxed AI agent workloads. You'll also own how we decide where workloads run, which is a real scheduling problem in our environment given the mix of GPU types, tenant isolation requirements, and capacity dynamics. Because our fleet spans real hardware in real facilities, the work reaches down into the physical layer too, and you'll be close to the deployment side rather than abstracted away from it. What you'll work on Building the virtualized compute lifecycle end to end, including provisioning, resize, live migration, and reclaim, exposed through a single API that upstream services can rely on regardless of what's running underneath Architecting a mixed hypervisor pool that handles both general-purpose tenant workloads and sandboxed environments for agentic AI, with routing logic that picks the right backend based on workload class Owning workload placement across the fleet, balancing capacity, isolation, hardware locality, and packing efficiency without letting the fleet drift into fragmentation over time Defining the API contracts and event streams that other services integrate against, including the hooks that feed usage tracking and billing Working with our hardware and deployment teams on node bring-up, firmware and BMC workflows, network fabric configuration, and the operational tooling that keeps sites healthy over time Partnering with the teams building on top of your layer to get the lifecycle semantics right, especially around failure modes and the edge cases that only surface with real traffic What we're looking for Strong systems background, comfortable with Linux virtualization, hypervisors, and the interfaces underneath them Hands-on experience with at least one modern virtualization stack, whether that's KVM-based, microVM-based, or a container runtime with real isolation, and interest in going deep on the others Working knowledge of the hardware layer, including server platforms, GPU and accelerator topology, PCIe and NVLink, high-speed networking, and the failure modes that show up once you're running real fleets Familiarity with deployment and provisioning workflows for bare metal, whether that's PXE and image-based flows, out-of-band management via IPMI or Redfish, or the tooling around firmware and driver lifecycle Good instincts for control-plane API design, the kind of API that other teams will build against for years and shouldn't have to fight Familiarity with scheduling and placement tradeoffs, particularly around multi-tenant isolation and heterogeneous hardware Comfort with event-driven architectures and message buses Clear technical writing, since a lot of the work here is specifying behavior and getting alignment before code lands Nice to have Experience running GPU infrastructure in production, especially for AI training or inference Background in multi-tenant cloud platforms, billing and metering systems, or telco-style OSS/BSS Exposure to NUMA-aware placement, live migration internals, or accelerator scheduling Hands-on time in data centers or edge sites, including rack-and-stack, cabling, and post-deployment validation Rust or Go for control-plane work, and comfort dropping into C when needed