Staff+ Software Engineer, ML Sampling Path

Anthropic•San Francisco, CA

1d•$320,000 - $485,000•Hybrid

About The Position

The Safeguards ML Sampling Path team builds and operates the production services that power Claude's safety systems. These services sit on the token generation path across every platform Claude runs on: every request must pass through them, and each millisecond of added latency is wait time for our users. You’ll keep p99 latency flat as traffic grows, build for robustness as dependencies time out or partially fail, and ship changes safely to a system that cannot go down.

Requirements

Designed, built, and operated high QPS systems at global scale, and were accountable for them in production: incident response, outages, and postmortem-driven remediation.
Strong foundation in distributed systems: replication, consistency tradeoffs, failure modes, and SLO management under load.
Design systems for graceful degradation: you plan for a slow dependency, a dropped stream, or a half-rolled-out deploy before it happens, and build so the system degrades predictably instead of failing.
Successfully shipped broad or all-encompassing changes to mission critical systems (e.g., database migrations, interface changes, rewrites).

Nice To Haves

8+ years of industry software engineering experience.
Familiarity with LLM inference systems and transformer-based models (not required, but a plus).

Responsibilities

Design, build, and operate the backend systems that process every token on the generation path for Claude requests, including the streaming contract with the API and inference engines.
Own latency and reliability end to end: define and maintain SLOs and error budgets for added latency, time-to-first-token, and availability, and lead incident response and postmortem follow-through.
Ship changes to the hot path rapidly but safely — canaried and gradual rollouts, error budget and latency gating, fast rollbacks — and drive per-token performance: chase tail latency and keep cost flat as traffic, models, and checks per request grow.
Set technical direction for the sampling path: lead design reviews, make latency, reliability, and cost trade-off calls with the inference and research teams, mentor engineers, and raise the operational bar for the wider Safeguards organization.