Your AI feature needs to survive production, not just demo day.

Problem

A job post now draws 200+ applications and most of them are AI-written. A small team without an HR function cannot read that pile. But a ranking they cannot explain is worse than no ranking at all, because it is a hiring decision nobody can defend afterwards.

Approach

The owner sets the rubric, up to twelve criteria, and candidates are scored against that rubric rather than the model's own idea of a good hire. Every criterion has to come back with a quote from the resume that earned it, enforced by the output schema, so a score with no evidence fails validation instead of reaching the user. A pile of up to a thousand resumes runs twelve at a time with per-candidate retries, so one malformed PDF is skipped and reported rather than killing the run, and progress streams while it works. A separate pass flags resumes that read as AI-written, and an eval suite with a fixed dataset keeps prompt changes honest.

Problem

Data products do not fail on the clever parts. They fail on the boring ones: the API that changes shape, the run that silently stops, the tab limit nobody planned for. Reliability is the feature.

Approach

Scheduled cron jobs pull three sports data sources every day, score batters over a rolling 14-day window and publish rankings to a Next.js dashboard. No server to babysit, and the job cleans up after itself so it never hits a storage limit.