JOB-1854

A small, early-stage AI company that builds custom language models for production agents needs an infrastructure engineer to take broad charge of how those models are trained and served. The hire works directly with the founding team and the research group.

Day to day

  • Take responsibility for the systems that train and host purpose-built language models, from research experiments to live traffic
  • Profile realistic workloads, then cut latency, raise throughput, improve GPU utilization and lower spend
  • Tune batching, caching, memory behavior and model execution, trying quantization or speculative decoding wherever they pay off
  • Provide inference capacity for evals, data creation and RL sampling, and build dependable model release and observability workflows
  • Help the founders and researchers divide GPU capacity across training and hosting, and get tuned models running inside customer setups

Qualifications

  • Practical history building or tuning LLM inference stacks, with a clear record of speed or stability gains
  • Solid systems engineering and troubleshooting instincts, including a good feel for what GPU memory and compute can handle
  • Experience with inference engines such as vLLM, SGLang or TensorRT-LLM, or similar tooling
  • Ability to balance output quality against speed, volume and spend, landing on practical answers
  • Comfort taking a prototype all the way into production on a small team
  • Bonus: CUDA or Triton kernels, inference spread across machines, serving several models at once, training infrastructure

On-site in the San Francisco Bay Area or the South Bay (Silicon Valley). Visa transfers considered case by case.