ML Systems Engineer (AI)

TL;DR

ML Systems Engineer (AI/Systems): Building high-performance systems for frontier model training and inference with an accent on RL feedback loops and hardware optimization. Focus on designing low-latency RDMA weight synchronization, optimizing GPU kernels, and orchestrating large-scale clusters to accelerate scientific discovery.

Location: Preferred Menlo Park or San Francisco, but can be flexible based on role

Compensation: $300,000–$400,000

Company

An AI and physical sciences company building state-of-the-art models to accelerate breakthroughs across materials, energy, and beyond.

What you will do

• Build rack and topology-aware scheduling for GB series GPUs across Ray, Slurm, and Kubernetes.
• Develop online/offline profilers to eliminate bottlenecks and implement direct S3 checkpoint streaming.
• Write and optimize communication and GPU kernels to maximize hardware throughput.
• Design and implement zero-copy RDMA weight synchronization between training and inference.
• Contribute upstream to open-source communities including SGLang, Megatron, and Ray.
• Co-design algorithms and infrastructure in close collaboration with RL and pretraining researchers.

Requirements

• Experience with large-scale inference infrastructure (load balancing, scheduling, serving architecture).
• Proficiency in low-level systems programming including RDMA, NVLink, and kernel-level optimization.
• Expertise in GPU cluster orchestration across Ray, Slurm, or Kubernetes.
• Ability to write and optimize CUDA kernels and distributed training collective operations.
• Experience profiling and benchmarking distributed ML systems across compute, memory, and network.
• Visa sponsorship is available for qualified candidates.

Culture & Benefits

• Opportunity to work on frontier-scale RL using thousands of GPUs.
• Direct impact on the pace of scientific discovery in physical sciences.
• High-ownership environment operating at the pace of the AI frontier.
• Legal and administrative support for visa sponsorship.