MLOps Engineer, LLM Performance (Kernels, Profiling, Serving)

ML systems expertise for AI evaluation

Help improve AI models’ understanding of the infrastructure used to train and serve large language models. The role centers on practical systems work in accelerator programming, performance profiling, distributed or accelerated workload debugging and inference delivery.

JAX, XLA and TPU experience is particularly valuable. The work involves writing and assessing difficult tasks and solutions that become training data for AI systems.

Technical responsibilities

• Design challenging problems across kernels, profiling, debugging and serving, including adversarial cases that expose model errors in accelerator code such as Pallas kernels on TPUs.
• Produce accurate, structured solutions and help researchers address gaps in systems, compilers and training infrastructure knowledge.
• Review technical tasks and answers, documenting feedback that can withstand specialist scrutiny.
• Develop scoring criteria for kernel optimization, profiler interpretation, distributed system reasoning and tradeoffs between serving throughput and latency.
• Coordinate with other experts to maintain consistent data quality.

Systems background

At least 2 years of direct professional work in ML systems, infrastructure, model serving or accelerator performance is required. The focus is systems engineering rather than applied modeling or data science.

Practical depth is needed in at least one area, with several preferred: custom kernels using Pallas, Triton, CUDA or CUTLASS; profiling with JAX or XLA tools, Kineto, torch.profiler or Nsight; distributed or accelerator constrained debugging; or large scale serving using vLLM, SGLang, TensorRT LLM, JetStream, KV caching, paged attention or continuous batching.

Frameworks and accelerators

Production experience with JAX or PyTorch is required. Additional framework depth may include custom operators, distributed training and sharding with pjit, shard_map, FSDP, Megatron or MaxText, or graph and compiler work with XLA, StableHLO, torch.compile or MLIR.

Experience with modern accelerators such as TPU generations v4 through v6e, H100, B200, Trainium or MI300 is expected, including judgment about memory, throughput and latency. Clear career progression and precise explanations of technical decisions are also important.

Engagement

The role is fully remote and open worldwide, with a 40 hour full time weekly commitment during weekdays. No other engagements or conflicting work commitments are permitted.

Employment or contracting is arranged through an employer of record partner appropriate to the contributor’s country, with payroll and any applicable benefits handled through that relationship. The assignment offers the opportunity to work within an AI laboratory’s extended workforce.