What the work actually is

This is not code review of model output. You are authoring the exam. For each task you pick or construct a realistic software engineering problem — a genuine defect, a missing feature, a performance bottleneck, a refactor that a legacy codebase badly needs — and package it as an environment a model can be dropped into. That means a pinned, reproducible repo state (Docker or equivalent), a deterministic test suite that passes only when the problem is genuinely solved, a golden reference patch you wrote yourself, and written reasoning explaining why the problem is hard and how a correct solution gets there. Tasks are authored in Python3, Java, Rust, Go, C++, or TypeScript — depth in one matters more than breadth across all six.

The hard part is usually the grading, not the fix. A test that passes on a hardcoded return value, a flaky timing-based performance assertion, or a bug whose fix is obvious from the stack trace all produce environments that teach a model nothing. Expect to spend real effort on adversarial self-review: what is the laziest patch that would make my tests green, and how do I close that hole? You will also review peer-submitted environments and reject ones that leak the answer or fail to rebuild cleanly.

What the screen looks for

micro1's screening is AI-led and follow-up heavy. It probes one language deeply rather than sampling all of them, asks you to walk through a specific production debugging episode in enough detail that invention collapses, and tests whether you can reason about test design and reproducibility — pinned dependencies, deterministic builds, isolation from network and wall-clock. No AI or ML background is expected or asked about; curiosity is enough. Vague seniority claims without a concrete repo, incident, or commit behind them are the main failure mode.

Logistics

  • Contractor, remote, fully async, roughly 15 hours a week with a weekly minimum task count.
  • Pay is output-based per accepted task; the $100–150/hr band is what contributors have reported as effective rates, not a guaranteed hourly wage. Rejected or reworked tasks eat into it.
  • Roles are typically filled within 48 hours, with first tasks expected 24–48 hours after onboarding. Apply when you can actually start.
  • You need your own development machine with Docker and the toolchain for your chosen language.