The opportunity
The Safeguards ML Infra team designs, builds, and operates the production infrastructure that powers Claude's safety systems. We own the critical backend services that ensure safety on the token generation path, and we own the operational work of getting those systems safely…
What you'll do
Launch captain model releases: stand up, configure, and verify safeguards for every new model, and serve as the safeguards point of contact in the launch room during release windows.
Own the off-cycle deployment of new safety classifiers as they ship from: research — canarying rollouts, running post-deploy validations, and investigating discrepancies when something looks wrong.
Verify that the right safeguards are provably live on the right models across: every deployment platform (1P, AWS Bedrock, GCP Vertex, etc.), and detect and eliminate configuration drift between them.
Automate yourself out of last quarter's work: turn launch runbooks into tooling, hand-built checks into continuous validation, and one-off deploys into a repeatable pipeline.
Plan to use Claude aggressively to do this! And be a trailblazer that paves: the path for safe agentic operations of safety-critical systems.
Build and maintain a safeguards registry with full provenance: what is running in production, on which model, on which platform, and when and by whom it was deployed.
What they're looking for
- Participate in on-call and operational-duty rotations covering service: incidents, model provisioning, and time-sensitive research and safety launches.
- Have owned production change management at scale: deploy pipelines, config management systems, canary analysis — and have strong opinions about what "verified" means.
- Have run high-stakes releases: served as a launch captain, incident commander, or release owner for systems where a bad deploy has real consequences, and are energized rather than drained by being in the critical path.
- Have meaningful on-call experience for production systems, including incident: response and postmortem-driven improvements — and a track record of turning (and fixing!) postmortem action items into process and tooling changes.
