A dated log of what I built, what broke, and what I finally understood. Written per milestone, unedited afterwards — the wrong turns are the useful part.
Xylodesk — Resume Studio, the résumé builder with templates and live editing.
GraphQL schema design, and testing enough to trust a refactor.
Shipping this site to GitHub Pages, then a proper accessibility pass.
Shipped skills and experience overlap scoring before touching Gemini. The AI tier now layers on top and falls back to the heuristic when a call fails. A working, explainable baseline beats a clever thing that sometimes returns nothing.
Switched acts_as_tenant to fail-closed. An unscoped query now raises instead of returning an empty set — because an empty set looks like "no results" in the UI and like a passing test in CI.
Spent the day building subagents that review my diffs rather than shipping features. Felt slow. It has since caught authorization gaps and tenant-isolation issues I personally wrote and personally approved.
Candidates can erase their data; employers must retain some of it. Landed on a 30-day window that anonymizes retained records in place, so foreign keys survive and the person does not. The hardest part was accepting that the naive answer was wrong.
A candidate profile and an employee HR record looked like one model for about a week. They are not: one is public, one holds compensation. Splitting them early cost an afternoon and removed an entire category of future leak.