What the work involves

You will be placed on one of two tracks after onboarding. On the connectors track, you build Python backend applications that faithfully replicate the behaviour of real SaaS products — Slack, Linear, Jira, Notion, Gmail, internal wikis — so that AI agents can be trained and evaluated against them without touching production systems. Fidelity is the whole point: matching endpoint semantics, pagination behaviour, permission errors, rate-limit responses and edge-case payloads, not just producing a happy-path REST stub. On the task track, you mine real data and workflows for long-horizon task candidates, author them as realistic scenarios, and write rubrics that draw explicit lines between correct, partially correct, and deficient agent work.

Both tracks expect heavy, fluent use of AI coding agents (Claude Code, Codex, Cursor, Copilot) across development, QA and validation — the project is partly a demonstration that experienced engineers can move fast with these tools while still catching what they get wrong. Work is unitised: a task is submitted, reviewed, and either approved or sent back. Pay accrues on approved tasks only, so rework is unpaid time.

What the screen actually measures

Expect probing on concrete backend specifics rather than résumé summary — which framework, which API you have emulated or integrated against, how you handled auth and pagination, what your test setup looked like. The screen also tests evaluation judgement: whether you can define a quality bar in writing that a second reviewer would apply the same way, and whether you can tell a genuinely hard long-horizon task from a long one. Availability is a real gate here, not a formality; the minimum is one approved task per day for four weeks.

Logistics

  • Remote, contractor, pay per approved task at an observed $300 per task
  • Four-week engagement starting immediately; rolling review of applications
  • Async work with scheduled onboarding, calibration and quality-review touchpoints
  • No cap on volume until the pipeline is exhausted; throughput is entirely on you
  • Own machine and environment: Git, Docker, and a working AI coding assistant subscription or access