Career Side Quests

• ✓It measures you, not your CV writing — most resume matchers score keyword overlap, which rewards whoever gamed the document best.
• ✓Gaps become scoped builds (side quests) and multi-hop routes (questlines), not a reading list.
• ✓It works at any distance — a sideways step or a full career change — instead of assuming you’re already nearly qualified.

Every job posting is a wall of requirements, and the honest question underneath it is uncomfortable: am I actually close to this, or am I kidding myself — and if I'm not close, what's the shortest way across? Nothing answered that. Resume matchers score keyword overlap, which really measures how well you wrote your CV. Asking a chatbot gives you a confident percentage that changes if you reword the question.

Both fail the same way: they produce a number nobody can check. I wanted the opposite — an assessment where every claim points at the exact line that justifies it, so a low score is something you can audit and argue with rather than just feel bad about.

The first version did one thing well: paste a role, upload a CV, and get back a requirement-by-requirement read. The pipeline was the product — pull requirements out of the posting, pull evidence out of your materials, then score one requirement per call with a skeptical reviewer. One contract that every stage speaks, and the aggregation deliberately kept out of the model's hands from day one.

• Evidence you can click on

  Every requirement gets a level and a verbatim quote from your own material — the exact line that earned it. The quote is checked as a literal substring of its source; if it isn’t there, the claim is thrown away and regenerated. Hallucination becomes a cheap string check instead of a leap of faith. “Your material” is deliberately wider than a CV: PDFs and Word documents, links, a GitHub profile treated as evidence rather than a decorative URL, and a box for the work no document records.

• Counter-evidence, on purpose

  Models flatter resumes by default — ask one to assess you and it will find a way to be encouraging. So each score is required to surface what cuts against it too, and the interface shows that alongside the supporting evidence. You get the case for and the case against.

• Side quests and questlines

  A gap isn’t a lecture, it’s a build. Each one routes to a side quest: one scoped project that closes it. When the target is a long way off, those chain into a questline — a multi-hop route across the distance, rather than pretending one weekend project bridges a career change.

• It reads a two-column CV properly

  PDF text extraction groups words by vertical position, which quietly splices a sidebar label into the middle of a job description whenever a wrapped line shares a line with it — “MBA” landing inside someone’s work history. Extraction now finds the column gutters first and reads inside each one, which is a large part of why the scores stopped moving between runs. Parse coverage is shown rather than swallowed, too — a CV that silently yielded one page out of fourteen would otherwise be scored as if that page were the whole person.

• Eligibility is a gate, not a score

  Averaging “can’t legally work here” into a capability percentage is wrong in both directions — it hides a blocker behind a decent-looking number, and it punishes a strong candidate for a fixable visa question. So eligibility sits outside the score as a gate.

This app tells people where they stand against something they want. That makes honesty a product requirement rather than a nicety — and almost every interesting decision came from enforcing it in code instead of trusting a model to behave.

Career Side Quests takes a role you want and whatever evidence you have, and hands back a requirement-by-requirement picture where every score cites the line that earned it, the counter-evidence is on the page, and each gap comes with the smallest build that closes it. It runs on my own infrastructure, deploys itself on every push, and treats being wrong loudly as better than being wrong quietly.

The number that matters most to me is repeatability: run the same CV against the same posting several times and the score used to swing 13 points. Column-aware PDF reading, a grounder that tolerates transcription drift but not fabrication, and a rubric rewritten as four ordered questions about what kind of evidence exists — rather than an adjective judgement — brought that down to 7, with ungrounded quotes going from 21 to zero.

It's still an alpha, and I'd rather say so. The pipeline works end to end and the honesty guarantees hold, but the rubric is still being tuned, the job-board adapters cover only some sites, and I'm still finding roles that score in ways I want to argue with. Try it on a real posting — just read the evidence, not only the number.