Building an ATS resume scorer

Can this document be read by machines, and is this candidate a good fit for this role? These have nothing to do with each other. A perfectly parseable CV can be a terrible match; a strong candidate can have a document that parses badly.

Blending them into one number is why candidates get advice that does not help. Report them separately: a structural score that is the same for every application, and a fit score that changes with the posting.

They also have different lifetimes, which is a practical reason to keep them apart. Parseability is fixed once and stays fixed; fit is recomputed per posting — so caching, cost and how often you show the result all differ between the two.

This half does not need a model. Extract text the way a parser would and inspect what came out: did sections come through in order, are dates readable, did the table in the header collapse into nonsense, is the contact information present.

Run the extraction and show the candidate what it produced. Seeing their own CV as a parser sees it is more persuasive than any score, and it turns abstract advice into an obvious fix.

Two-column layouts cause more of these failures than everything else combined. Text extracts in the document’s internal order rather than the visual one, so a job title from the sidebar lands under the wrong employer and the result is structurally valid and completely wrong.

• Text extraction succeeds and preserves reading order
• Sections identifiable — experience, education, skills
• Dates in a consistently parseable form
• Contact details present and extractable
• No content trapped in images, text boxes or multi-column layouts

Scoring a CV without a job description produces generic advice, because fit is only meaningful relative to a specific role. A tool that scores a CV alone is measuring conformity to a template.

With a posting, extract its requirements and check each against evidence in the CV. Report per requirement — met, partially met, absent — because that is what the candidate can act on, and it is what a single percentage destroys.

Weight the requirements rather than counting them. Missing one thing named in the job title is a different situation from missing three items in a list of twelve nice-to-haves, and a system that treats those as the same number is giving the same advice to two candidates in entirely different positions.

A score of 73% implies a measurement that does not exist. Nobody knows what the employer’s threshold is, how many others applied or how their system weighs anything, and presenting a precise number claims knowledge you do not have.

Bands are more honest and just as useful: strong, worth applying, likely to be screened out, missing a hard requirement. Candidates make the same decision from a band, without over-trusting a number that was never measured.

Precision also invites the wrong behaviour. A candidate shown 73% will try to reach 80, and the fastest route is padding the CV with the posting’s vocabulary — which improves the number, makes the document worse to a human reader, and is exactly what the tool should not be encouraging.

Keyword stuffing is the obvious one. Telling a candidate to insert terms they cannot evidence produces a document that passes a filter and fails the interview, and the tool is optimising the stage it can see rather than the outcome the candidate wants.

Design conformity is the quieter one. Advice to remove all formatting, adopt one template and match a standard structure produces documents that parse cleanly and are indistinguishable from each other, which helps nobody once every candidate has taken it.

And avoid inventing thresholds. "ATS systems reject CVs under 60%" is a claim with no source, and stating it confidently makes a candidate act on something invented — which is worse than saying that the employer’s actual criteria are not knowable from here.

• Never suggest a term the candidate cannot evidence
• Never claim a specific employer threshold
• Never recommend formatting that removes a document’s readability for humans
• Never present a fit judgement as a measurement

A finding the candidate cannot act on is noise. "Your CV scores 62" changes nothing. "This posting asks for Kubernetes and your CV never mentions it — if you have used it, say where" changes the next version.

Sort findings by how much they would move the outcome, and be explicit when the answer is that the CV cannot fix it. Telling someone the requirement is genuinely absent is more respectful than implying a rewrite would help.

Show three findings rather than fifteen. A long list is read as a verdict on the candidate rather than a set of edits, and the three that matter get lost among twelve that do not.