Container runtimes were designed for general-purpose software workloads. AI inference is not a general-purpose workload
Running large models at production scale exposes cracks in every layer of the container stack: runtimes unaware of GPU memory constraints, images that take minutes to pull when a model needs to scale to thousands of replicas, and isolation mechanisms that weren't designed for the multi-tenant serving environments that production AI requires. The tools the industry has relied on for a decade weren't built for this, and patching around those limitations at higher layers only goes so far
Baseten owns the entire pipeline, from the moment a developer pushes a model to the moment a request gets a response. That vertical ownership means we can fix these problems at the root. The Runtime Fabrics team is doing exactly that: purpose-building the container runtime and storage layers for AI inference workloads, led by some of the world's top containerd maintainers
As Engineering Manager of the Runtime Fabrics team, you will lead this work, setting technical direction, growing a world-class team of systems engineers, and ensuring the team's output shapes not just Baseten's infrastructure but the open-source container ecosystem at large. If you've contributed to containerd, runc, or related OCI projects and are ready to lead a team solving some of the hardest problems in infrastructure today, we'd love to talk
Guide the team in extending and hardening containerd, runc, and related OCI ecosystem projects to meet the GPU-specific requirements of production AI inference, including startup performance, GPU device access, and multi-tenant isolation
Oversee the architecture and evolution of the Baseten Delivery Network: the tiered caching and weight delivery system that makes cold starts 2–3x faster and eliminates thundering herd failures during burst scaling events
Drive the expansion of BDN's architecture, currently focused on model weights, to container images, training checkpoints, and deployment artifacts
Provide technical oversight on GPU-aware isolation mechanisms for multi-tenant inference, including secure container runtimes, Linux namespace hardening, and longer-term micro-VM integration
Ensure the team maintains end-to-end ownership of the container startup performance path, from snapshotter initialization through weight delivery to first inference request
Champion the team's contributions back to the open-source containerd ecosystem alongside a team of core maintainers