Recruiter: see how your CV matches a job post

The problem

A Danish job post asks for six or eight things, written as sentences: “uddannet pædagog eller pædagogisk assistent”, “flydende dansk”, “gerne erfaring med Kubernetes”. Some are must-haves, some are nice to have, and some are alternatives. A candidate reads the post, then their own CV, and guesses.

Keyword matching, the usual answer, gets this wrong in both directions. It counts “Kubernets” with a typo as missing, and it cannot tell that a CV naming a closely related skill is partly there. I spent years on a job and candidate matching platform at JUICE, and the ranking engine I built there had the same blind spot: a requirement was either in or out.

What cv_match does

cv_match is the core of Recruiter. You upload or paste a CV, then paste a job post, upload one, or give it a link. You can add up to ten posts and compare them against the same CV.

For each post it shows:

• every requirement, marked met, partly met or missing;
• the phrase from your CV that decided it, and a one-line reason (“Needs 3, CV shows 5.”, “Named in the CV.”);
• the soft skills as tags, not scored, because no CV proves “engageret”;
• one match score, with the posts ranked by it.

From a match it writes a job application and a résumé tailored to that post, both from your own CV, with room for notes on what to stress. Matches are saved, so you can come back to them.

How it decides

Each requirement goes through the cheapest step that can decide it, and only what is left goes to a language model:

1. Rules. Years of experience are arithmetic: a post asking for 3 years against a CV showing 5 is met, against 2 it is partly met.
2. The taxonomy. A Danish and English vocabulary of 9,926 skills, 401 roles and 620 job titles, with 23,533 labels and synonyms. A requirement the CV names, spelled the same or with a typo or two, is met. One that names a broader, narrower or sibling skill is partly met.
3. The model. Requirements written as sentences, levels (“flydende dansk”, a master’s degree) and responsibilities go to the model, which answers with a verdict, the evidence and a reason.

Alternatives (“pædagog eller pædagogisk assistent”) count as one requirement, met by the best of them. The score weighs formal requirements at 150, nice-to-haves at 50, and the job title and responsibilities at 30 each; a partial match counts half.

Pasted links are read carefully: only public web pages, every redirect checked, and the post’s own structured data when the job board has it. Page fetches are capped per session, so a list of links can’t turn it into a crawler.

Free, or with your own key

Without an account you get the free keyword match: the taxonomy and the rules, no model, nothing that costs anything per request. It is honest about what it can’t do. It finds requirements the post names as a skill, tool or title, and counts anything your CV doesn’t name word for word as missing.

Sign in and add your own OpenAI or Anthropic API key, and you get the AI match: it reads the whole post, judges related experience, fetches posts from a link, saves your matches and writes the application. The model calls run on your key, so Recruiter never needs to charge for them.

How accurate it is

Before building the match, I measured the step that turns a job post into a list of requirements. 200 real Danish job ads, with expected fields labelled one ad at a time and spot-checked by hand:

“Hard fields” are the ones a default can’t get: title, education level, years, skills and certifications. The WebLLM models I tried in the browser scored between 17% and 57% and needed a 1.5 to 3 GB download, so extraction runs on the server. The postcode and full or part time come from plain rules, which matched the labels 98% and 82% of the time.

For a second opinion, a match can also go to Laya, a small decision model. Given the job post and CV raw, in Danish, it was close to chance. Given the code’s own verdicts as one English sentence (“meets 3 of 5; missing: …”), it separated candidates who meet every requirement from those who don’t with an AUC of 0.98. That is on 20 hand-written pairs, so its thresholds still need calibrating against real recruiter decisions.

Built on trongate.cloud

Recruiter is a Trongate app, and it runs on trongate.cloud, the same way a customer’s app does: push to GitHub, and it is built and deployed to recruiter.trongate.dev with its own database. That is the point. A hosting platform is only proven by running something real on it, and Recruiter is the app I would have built anyway.

Try it

Match your CV at recruiter.trongate.dev.