Staff ML Systems Engineer, Sampling Path (Safeguards)

EngineersOfAI; San Francisco, Northern (CA, KY); Hybrid; USD 320,000 - 485,000; Full time

Benefits offered by this job

Visa sponsorship

Job summary

Anthropic is seeking a seasoned software engineer for the Safeguards ML Sampling Path team in California. You will design and operate backend services powering Claude’s safety systems, ensuring low latency and high reliability across the token generation path.

You will own SLOs, respond to incidents, and implement safe deployment strategies while collaborating with inference and research teams to optimize per-token performance and cost as models grow.

Qualifications

• 8+ years of industry software engineering experience.
• Familiarity with LLM inference systems and transformer-based models (not required, but a plus).
• Designed, built, and operated high QPS systems in production with incident response and postmortem follow-through.

Responsibilities

• Design, build, and operate 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, lead incident response and postmortem follow-through.
• Ship changes to the hot path rapidly but safely with canaries, gradual rollouts, and latency gating; drive per-token performance and cost control.
• Set technical direction for the sampling path; lead design reviews, balance latency, reliability, and cost with teams, mentor engineers.

Skills

High QPS systems

Distributed systems

Latency optimization

Incident response

Education

Bachelor's degree in computer science or related field

Job description

Anthropic is seeking a seasoned software engineer for the Safeguards ML Sampling Path team in California. You will design and operate backend services powering Claude’s safety systems, ensuring low latency and high reliability across the token generation path.

You will own SLOs, respond to incidents, and implement safe deployment strategies while collaborating with inference and research teams to optimize per-token performance and cost as models grow.