Staff+ Software Engineer, Inference Runtime
AnthropicAbout Anthropic
Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems.
About the role
Anthropic's Inference organization serves Claude to millions of users and enterprise customers with the speed, reliability, and efficiency that frontier AI demands. We build across GPUs, TPUs, and Trainium, and the complexity of our development environment grows with every platform we add.
We're looking for a Staff Engineer to be a technical lead for Inference Runtime: the team that owns the shared, accelerator-agnostic core of our inference serving stack, whose performance, correctness, and abstractions every accelerator builds on.
This is a senior IC role with broad technical ownership. You'll set technical direction for the runtime's architecture, its release and validation systems, and the workflows engineers use to develop on top of it. You will partner across Inferencing to make hard calls on boundaries, prioritization, and tradeoffs across heterogeneous accelerator platforms.
You'll pair with the team's Engineering Manager, who owns hiring and people development, while you own the technical roadmap and drive the work, representing the team in cross-org efforts spanning serving, scaling, and accelerator teams.
This role is for someone who has been the technical anchor of a platform with many internal consumers, who thinks in systems and feedback loops, and who gets real satisfaction from building abstractions that hold up as the system scales another order of magnitude.
Key responsibilities
-
Set technical direction for the team, owning the architecture and roadmap for the shared runtime of the inference serving stack
-
Own and evolve the accelerator-agnostic runtime itself – its interfaces, internal boundaries, and build structure – including hands-on work in a performance-sensitive Rust and Python codebase
-
Keep the platform's expansion cost low by ensuring new models and deployment targets pay only for their own specialization, and edge cases stitch back into the core easily
-
Drive efficient accelerator usage – utilization, scheduling, memory management – across GPU, TPU, and Trainium
-
Build the runtime's validation surface around partitioned builds, change-scoped t