Staff + Senior Software Engineer, Inference Deployment

About the Role

Our mandate is to make inference deployment boring and unattended.

Anthropic serves Claude to millions of users across GPUs, TPUs, and Trainium — and every model update must reach production safely, quickly, and without disrupting service. We're building the systems that make inference deployment continuous and unattended.

As a Software Engineer on the Launch Engineering team, you'll design and build the deployment infrastructure that moves inference code from merge to production. This is a resource-constrained optimization problem at its core: validation and deployment consume the same accelerator chips that serve customer traffic — your deploys compete with live user requests for the same hardware. Every model brings different fleet sizes, startup times, and correctness requirements, so the system must adapt continuously. You'll build systems that navigate these constraints — orchestrating validation, scheduling deployments intelligently, and driving down cycle time from merge to production.

If you've built deployment systems at scale and gravitate toward the hardest problems at the intersection of automation and resource management, this team will give you an outsized scope to work on them.

Responsibilities

• Own deployment orchestration that continuously moves validated inference builds into production across GPU, TPU, and Trainium fleets, unattended under normal conditions
• Improve capacity-aware deployment scheduling to maximize deployment throughput against constrained accelerator budgets and variable fleet sizes
• Extend deployment observability — dashboards and tooling that answer "what code is running in production," "where is my commit," and "what validation passed for this deploy"
• Drive down cycle time from code merge to production with pipeline architectures that minimize serial dependencies and maximize parallelism
• Optimize fleet rollout strategies for large-scale deployments across thousands of GPU, TPU, and Trainium chips, minimizing disruption to serving capacity
• Evolve self-service model onboarding so that new m