Staff ML Inference Systems Engineer, Real-Time Safety

EngineersOfAI; San Francisco, Northern (CA, KY); Hybrid; USD 272,000 - 368,000; Full time

Job summary

Anthropic in San Francisco, CA, seeks an experienced ML infrastructure engineer for the Safeguards Inference Path. You will design, build, and run production ML systems powering Claude's safety classifiers, collaborating with researchers and inference engineers to translate safety research into scalable production.

You will work at the intersection of ML, distributed systems, and safety, delivering robust tooling and platforms that support model launches at scale with reliability and low latency.

Qualifications

• Proficient in Python with strong ML tooling experience.
• Experience building scalable ML infrastructure for production.
• Hands-on with ML frameworks (PyTorch, TensorFlow, JAX).
• Familiarity with real-time, low-latency systems and monitoring.
• Ability to collaborate across research, safety, and engineering teams.

Responsibilities

• Design and build scalable ML infrastructure for real-time safety deployments.
• Develop monitoring and observability tools for classifier performance.
• Collaborate with researchers to productionize safety research.
• Optimize latency and throughput for real-time safety evaluations.
• Implement automated testing, deployment, and rollback for ML models.
• Partner with safety, security, and alignment teams on requirements.

Skills

Python

Distributed systems

A/B testing

Collaboration

Tools

PyTorch

TensorFlow

JAX

Job description

Anthropic in San Francisco, CA, seeks an experienced ML infrastructure engineer for the Safeguards Inference Path. You will design, build, and run production ML systems powering Claude's safety classifiers, collaborating with researchers and inference engineers to translate safety research into scalable production.

You will work at the intersection of ML, distributed systems, and safety, delivering robust tooling and platforms that support model launches at scale with reliability and low latency.