What the work involves
You write, profile, and optimize CUDA kernels — and then explain your reasoning in a form a model can learn from. In practice that means taking a kernel or a C++/CUDA codebase, identifying where throughput is actually lost (occupancy, memory coalescing, bank conflicts, register pressure, launch overhead), applying targeted changes, and documenting the before/after with numbers from Nsight Compute, Nsight Systems, or equivalent tooling. Some tasks extend into graphics and compute shaders via GLSL or WebGPU, where the same discipline applies to a different runtime.
A meaningful share of the work is evaluative rather than generative: reading model-produced CUDA, deciding whether an "optimization" is real or cargo-cult, and writing critiques that hold up under review. Prior AI or ML experience is not required — the value you bring is knowing why a kernel is slow on an A100 but fine on an H100, and being able to say so precisely.
What the screening looks for
- Concrete evidence of performance-tuning work: specific kernels, specific speedups, specific hardware generations.
- Fluency under follow-up questioning — vague answers about "using shared memory" don't survive a second probe.
- Written clarity. Reports and rationales are the deliverable, not just the code.
- Honest calibration: knowing where your expertise ends is scored, not penalized.
Logistics
Remote, contractor, largely asynchronous with occasional synchronous design discussions. Most contributors work 10–20 hours per week on a rolling task queue, though allocations vary with customer demand. Pay is stated as observed for this listing and depends on assessed depth and task type; nothing is guaranteed.