Machine Learning Infrastructure Engineer

About the Role

Join an AI team developing foundation models for physical systems, with weather as an initial application. You will build the distributed infrastructure that supports large-scale model training and inference, helping advance AI for science and physical AI.

What You'll Do

• Design, deploy, and maintain distributed machine learning training and inference clusters.

• Build scalable end-to-end pipelines for petabyte-scale datasets and the full model lifecycle.

• Research and evaluate training approaches, including parallelization strategies and numerical precision trade-offs across model scales.

• Profile and debug low-level GPU operations to improve performance.

• Analyze experiment results, including through ablation studies, and apply relevant research to infrastructure work.

What We're Looking For

• Two to ten years of experience building large-scale machine learning infrastructure for pre-trained or foundation models.

• Hands-on experience across compute clusters, data infrastructure, distributed training, and inference, with a record of optimizing large-scale workloads.

• Proficiency with distributed training frameworks such as FSDP or DeepSpeed, and experience with GPU performance optimization and debugging, including CUDA and JAX.

• Knowledge of cloud platforms such as GCP, AWS, or Azure, plus containers and orchestration tools such as Docker and Kubernetes.

• Familiarity with model serving, distributed task management, and ML system monitoring, logging, observability, and version control.

• Experience in AI for science or physical AI, such as robotics, autonomous systems, or biology, and a generalist approach across the ML lifecycle.

• Comfort working in fast-paced, time-sensitive environments and learning unfamiliar domains.

Compensation & Benefits

Base salary: $200,000 to $400,000 annually. Visa sponsorship is not available.

Location

On-site in San Francisco, California, United States, five days per week. Relocation support is provided.