SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Salary: USD 320,000 - 485,000 / annual
Anthropic is seeking a Staff Software Engineer to join the Safeguards ML Inference Path team, which designs, builds, and operates production infrastructure for Claude's ML-based safety systems. This is a high-impact role at the intersection of machine learning, distributed systems, and AI safety.
You will own the research-to-production pipeline for safety technologies, working on the critical path for every Claude model launch. Responsibilities include designing and building scalable ML infrastructure for real-time safety deployments across classifiers and models; creating monitoring and observability tools to track classifier performance and system health; collaborating with research teams to productionize safety research; optimizing inference latency and throughput for real-time safety evaluations; implementing automated testing, deployment, and rollback systems for ML models; and partnering with Safeguards, Security, and Alignment teams to deliver infrastructure meeting safety and production needs.
Ideal candidates have deep expertise in productionizing ML systems, proficiency in Python and ML frameworks (PyTorch, TensorFlow, JAX), understanding of distributed systems principles, and experience building high-throughput, low-latency systems. Strong candidates typically have 5+ years building production ML infrastructure in safety-critical domains like fraud detection, content moderation, or risk assessment; experience with large language models and transformer architectures; expertise in monitoring and alerting for ML model performance; and knowledge of privacy-preserving ML techniques.
You should be results-oriented with a bias toward reliability in safety-critical systems, enjoy collaborating with researchers to translate cutting-edge research into production, and care deeply about AI safety and societal impact. The team serves thousands of ML classifiers for all token generation requests across all Claude platforms (1P, Bedrock, Vertex, and beyond), handling increasingly complex and frequent model launches.
