ML Systems Engineer, GPU Performance (Seed-Stage AI Infrastructure)

A seed-stage startup is rebuilding the software that trains and serves generative world models, working from GPU kernels all the way up to multi-node serving. The engineer joining owns how fast and how efficiently the whole stack runs, alongside founders who bring experience in distributed systems, kernels, cloud and research.

Your scope

• Tune GPU and system performance for training and inference of imagery, video and world models
• Hunt down bottlenecks across kernels, memory, hosts and clusters with profilers such as Nsight
• Hand-write performance-critical code in CUDA and Triton
• Create and refine distributed engines for serving and training diffusion models
• Own speed of communication between GPUs and nodes, covering NCCL, RDMA over InfiniBand or RoCE, and split-stage (disaggregated) inference serving
• Create benchmark and regression suites that keep speedups from eroding in production
• Co-design kernels, runtime and models with the underlying hardware in mind

Who fits

• At least two years working on deep learning inference or training systems and distributed systems
• Authored core functionality inside an inference or training framework such as vLLM, SGLang, TensorRT-LLM or Megatron
• Hands-on kernel development on NVIDIA GPUs with CUDA, CUTLASS, Triton or PTX
• Strong ML and computer science fundamentals, with a degree in CS or a related quantitative discipline
• Optimizing diffusion, video, image or other multimodal models is a plus

On-site in the South Bay Area of Silicon Valley. Visa transfers and new sponsorships are both possible.