Clearing the filters without inventing you

After enough applications, it gets tempting to let the résumé drift a little further from the truth each time, because the posting wants "Kubernetes in production" and what you actually have is a homelab and a strong opinion. A screen may not catch the difference. An interviewer can.

I don't think people do this because they're dishonest. They do it because tailoring a résumé honestly requires source material, and a résumé is terrible source material. It's a compressed summary you wrote years ago, and when a posting demands different emphasis there's nothing underneath it to re-draw from. So the writer reaches for the only thing in reach, which is imagination.

That's the actual problem jobhunt was built around, and the fix is unglamorous: get a rich, true record on file before any tailoring happens. Coach interviews you in chat, one question at a time, and builds a career profile out of your answers. What you built. What was yours versus the team's. The real numbers, hedges included. The things you never want claimed on your behalf. Slow compared to "upload résumé, click generate"? A bit. It gives the downstream tools real material to work with instead of asking a model to fill the gaps.

Once that record exists, tailoring does not need to be a creative writing exercise. The system reads the posting for the terms that actually matter and organises your evidence around them. Where your record covers a requirement, the résumé can say so in the posting's own language. Where it doesn't, the gap belongs in your fit analysis instead of a bullet point. Generation can still fail at that boundary, which is why the validation pass exists.

A few implementation details do more work than you'd expect.

The posting itself is stored verbatim and analysed as-is. Early on we could have let the model summarize job descriptions before matching against them, and it would have been cheaper. We didn't, because paraphrase is where requirements get quietly rounded toward whatever the model finds convenient. If the posting says "5+ years of Go," the analysis should be looking at those words, not at a summary that softened them to "backend experience."

Your name, email, and phone number are stamped into documents by the server, from your saved details. The model literally cannot touch them. This sounds paranoid until you've watched a language model "fix" a phone number.

And after a résumé or cover letter is generated, a separate pass audits it, claim by claim, against your profile and the posting. We added that pass after our document evals found claims that couldn't be traced to anything the person had said; I've published the numbers elsewhere. The flags show up next to the download. Nothing gets silently rewritten, because a tool that edits your claims behind your back hasn't solved the honesty problem, it's just moved it somewhere you can't see.

The result isn't magic. It may be a one- or two-page PDF, depending on how much material has to remain readable. A separate validation report says whether the audit passed and lists claims it could not trace. A flag is not silently removed, and the document is still available, so you should resolve it before using the file.

The filters were never one enemy with one score. Relevance can be established honestly if the truth about your career is written down somewhere in enough detail. A tool that skips that step can only generate on top of a thin source. This one asks you to spend time on the record first. I think that's the right trade, but then, I would.