What the work actually is

You are not shipping a serving stack. You are constructing the problems that reveal whether a frontier model actually understands one — and then judging its answers. A typical unit of work starts with a scenario you draw from real experience: a paged-attention KV cache eviction policy that thrashes under bursty long-context traffic, a fused softmax kernel that loses numerical stability at fp8, a continuous-batching scheduler whose tail latency collapses when prefill and decode contend for the same SMs. You write the scenario so it has a defensible correct answer, author the reference solution at the level a strong senior engineer would produce, and build a rubric that separates a merely plausible answer from a correct one.

Then you grade. Model outputs in this domain are often fluent and wrong in interesting ways — correct CUDA syntax with a race on shared memory, a roofline argument that cites the wrong bottleneck, a vLLM configuration recommendation that would be right for throughput and wrong for the latency SLO stated in the prompt. Your job is to catch that precisely and write down why, in prose another reviewer can audit.

What the screen looks for

  • Hands-on recency. Named systems, kernels you have profiled, versions, hardware. "Worked on inference" does not survive a follow-up.
  • Depth under pressure. The screener will push one or two layers past your first answer — into memory hierarchy, occupancy tradeoffs, quantization error behaviour, or scheduler design.
  • Rubric instinct. Can you articulate the difference between an answer that is wrong and an answer that is right for the wrong reason, and does your grading distinguish them?
  • Writing. Structured feedback is the deliverable. Terse, precise technical prose beats volume.

Logistics

Fully remote and asynchronous — no standing meetings, work claimed and returned on your own schedule. Contributors commonly run 5–15 hours a week alongside a full-time role, with the practical constraint being consistent weekly throughput rather than fixed hours. Pay is stated as $100–170/hr and scales with track depth and demonstrated evaluation quality; the band is what the platform advertises, not a guarantee for any given contributor.