What the work actually is

You write tasks that a strong model should be able to do and often can't — then you prove which is which. A representative item might ask for a corrected implementation of an attention variant, a reproducible fine-tuning workflow that fits a stated memory budget, a fix for a silent dtype-induced numerical drift, or a throughput optimization on a vLLM or SGLang serving path. For each one you supply a reference solution, objective tests or benchmarks that pass only for correct implementations, and a written account of the trade-offs and failure modes involved. You will also review AI-generated code and other contributors' submissions: deciding whether an output is genuinely correct, merely plausible, or correct-but-unreproducible is a large share of the job.

What the screen is looking for

The ~30-minute AI interview probes below the API surface. Expect follow-ups on things you claim: which framework you used, what you personally changed, how you measured the improvement, and what broke first. Generic familiarity with PyTorch does not survive a second question about autograd behaviour, memory layout, or why a distributed run hung. micro1 asks for a master's or PhD in a quantitative discipline plus real experience with at least two tools from the modern ML stack — PyTorch, JAX, NumPy/SciPy, Hugging Face Transformers or Tokenizers, vLLM, SGLang, llama.cpp, or defensible equivalents. A technical assessment may follow, then hiring manager review.

Logistics

  • Fully remote, global, contractor engagement at roughly 15 hours per week
  • Hours are self-scheduled, weekends included; there are no standing meetings
  • Compensation is output-based — paid per accepted task against project spec, so the observed hourly band depends on how fast you produce work that passes review
  • Minimum submission requirements apply, and roles fill fast: first task is typically expected within 24–48 hours of onboarding

This suits people who like precise, self-contained technical problems and can defend an implementation choice in writing. It suits people who want steady ticketed engineering work less well.