What the work actually is
You write the problems, then you judge the machine's answers. A task might be a broken GitHub Actions matrix build, a Terraform module that provisions an unreachable service, a Kubernetes deployment that passes readiness but drops traffic on rollout, or a postmortem-style incident scenario. For each one you supply a reference solution that actually runs, with setup steps and validation commands someone else can reproduce on a clean machine. Then you review model output: locate the failure precisely, correct it, and write the reasoning that explains why the model went wrong — not just that it did.
The grading half is where most of the difficulty sits. Model output in this domain is usually plausible-looking YAML or HCL that is subtly wrong: a container that starts but has no liveness probe, an IAM policy far broader than the task needed, a pipeline that appears green because the test step never executed. Your job is to catch that and articulate it in a way a downstream trainer can use.
What the screen looks for
micro1's screening is AI-led and includes a live technical conversation with follow-ups that go deeper than your first answer. It is looking for hands-on production experience rather than certifications — outages you actually handled, pipelines you actually owned, state files you actually recovered. Real coding ability matters here: the listing is explicit that Bash glue alone is not enough, so expect to discuss Python, Go, or TypeScript you have written. Written English is assessed directly, because corrections and rationales are the deliverable.
Logistics
- Contractor engagement, fully remote, no fixed office hours
- Asynchronous task and review work with occasional calibration threads or calls
- 20 hours/week minimum; many contributors treat it as a defined part-time block
- No prior AI or ML background required — domain depth is what is being bought
- Expect a calibration period where early submissions get heavy reviewer feedback before volume opens up