What the work actually is

You write analytics tasks that look like real jobs, not textbook exercises. That means starting from raw, slightly messy source tables — late-arriving rows, duplicated keys, timezone drift, a slowly changing dimension nobody documented — and building through to a tested model with a reference answer. Alongside the answer you write the reasoning: why a window function instead of a self-join, why this grain, why the incremental strategy, what breaks if the model picks the obvious path instead.

The second half of the job is grading. You read model-generated SQL and decide whether it is right, and separately whether it is sane. A query can return correct numbers and still scan two terabytes because it filtered after the join or ignored the partition column. A query can also read beautifully and be silently wrong — fan-out from a one-to-many join, a `LEFT JOIN` filter pushed into the `WHERE` clause, `COUNT(DISTINCT)` where the grain was already unique. Catching the second category under time pressure is the skill the labs are paying for.

What the screen looks for

AfterQuery's screening is AI-led and follow-up heavy. It probes concrete warehouse experience — which engine, at what data volume, what you did when a model got slow or expensive — rather than tool checklists. Expect to be asked to reason aloud about a query with a subtle defect, and to justify a modeling decision that has a defensible alternative. Written English matters more than usual here, because the reasoning you write is the deliverable, not a byproduct.

Logistics

  • Fully remote, fully asynchronous; no standing meetings and no fixed hours.
  • Minimum commitment of 10 hours per week, set by you; work is ongoing rather than a single batch.
  • Paid weekly via Stripe. The $60–100/hr band is as observed on the platform and typically tracks task complexity and review track record rather than being negotiated up front.
  • Expect a paid or unpaid sample task as part of onboarding — usually authoring one scenario end to end plus grading a handful of model responses.