GPU Kernel and ML Systems Researcher (Research Coding Tasks)

What You'll Do9

• Pick a recent paper with released code, ideally one you wrote, and isolate the implementation that delivers its central result.
• Remove that core code and wrap the remaining repository into a self-contained task an AI agent can attempt.
• Write the task statement so an agent understands what to rebuild and which benchmarks count.
• Define held-out evaluation workloads and set pass thresholds for each metric.
• Build a hidden reference solution that restores the paper's contribution.
• Add a second solution that clears the thresholds, proving the bar is reachable.
• Confirm the task reproduces the authors' reported improvements over the baseline.
• Review agent submissions against your task and judge how each attempt performs.
• Keep the environment reproducible so results hold across runs.

Requirements8

• Published research in ML systems or GPU kernels at venues such as MLSys, OSDI, SOSP, PPoPP or SC, or equivalent work.
• Code released alongside at least one of your papers, with a GitHub repository or artifact link you can share.
• PhD in computer science or a related field, either completed or in progress.
• Hands-on depth in CUDA, Triton or CUTLASS kernel development.
• Familiarity with inference serving and ML compiler stacks.
• Comfort with Docker and benchmark setups that others can reproduce.
• A first-author paper that reports measured gains over a baseline stands out.
• Ability to write clear task specs and judge technical work against stated metrics.

Who Should Apply

PhD researchers and research engineers who publish in ML systems or GPU kernels and ship code with their papers. You have a first-author result you can point to, a repository to match, and the patience to turn that work into a benchmark other people can run.

Salary Insight