Software Engineer, AI Tooling

MatX•Mountain View, CA

1d•$120,000 - $600,000•Hybrid

About The Position

MatX builds custom silicon for training and running the world's largest AI models, with hardware/software co-design across ISA, RTL, simulator, compiler, and kernels. The AI tooling team sits across that whole stack and owns the agent harnesses and shared workflows that turn general-purpose coding agents into ones maximally useful on each layer. Every engineering team using coding agents sits somewhere between watching every command and waking up to agent-written PRs waiting for review. What moves teams along this trust curve is how far engineers can trust their agents' judgment. An agent's judgment is only as good as its context and its ability to verify its work. However, context can be spread all over (think repos, design docs, Slack threads); the builds and simulations an agent needs to check its work may only be available on machines it can’t reach. You'd build infrastructure to address these gaps, at a company small enough your work reaches every engineer.

Requirements

Have run production infrastructure somewhere dynamic and know which shortcuts you’ll regret.
Use coding agents heavily and have strong opinions on what you let them do, what you don't, and how you verify their work.
Built engineering workflows (methodology, flow automation, review tooling, agent harnesses) that other engineers adopted and kept.
Notice LLM agents asserting something confidently wrong and modify their harness in response.

Nice To Haves

Have operated Bazel remote execution or a remote cache.
Are familiar with sandboxing and ephemeral compute for untrusted workloads.

Responsibilities

Make agents legible, first-class users of our infrastructure. Each one gets an identity of its own, reaches exactly what its task needs and nothing else, and leaves a trail a person can later audit.
Build environments for our agents to run in, e.g. short-lived VMs that exist for one task and no-egress sandboxes for long runs.
Build our company agent knowledge base, the context agents start tasks with and the place they go to get more. Pull together information spread across our monorepo, design docs, and Slack and structure it for agent consumption.