Sebastian Bochenek

I built Keywordise. Most of my work is the unglamorous half of hiring software: what happens to a CV between the moment it is uploaded and the moment a recruiter reads a row in a database.

What I actually do

Keywordise pools job postings from European employers and files applications through the employers' own hiring software. Building that meant reading how those systems ingest a document, which is how this blog started: almost everything written about ATS-friendly CVs is repeated advice with no measurement behind it, and a good deal of it is wrong.

I do not sell CV templates and I do not run a recruitment agency. That matters when you read the recommendations here, because the usual incentive to tell you your CV needs rewriting is not present.

How the numbers in these articles are produced

Every figure published here comes from running real documents through text extraction and counting what survives. It is not an opinion poll and it is not vendor marketing.

• Extraction. Documents are parsed with pdfminer.six in two modes: a column-aware layout pass, and a strict position-ordered pass that reads the page the way a naive parser does. A CV that only survives the first mode is a CV that depends on luck.
• Scoring. Nine checks, and the code is the specification: the name is the first line; the email survives as one token; a phone number survives as one token; each of the three section headings appears as a literal word rather than letter-spaced; the contact block is not interleaved into a job entry; no ligature glyphs corrupt a keyword; and no orphan bullet characters are left on their own lines.
• Reporting. Failures are published alongside passes. The two-column CV that scored four of nine in the parser comparison was my own, and it failed on the name, on all three headings, and on 45 orphan bullet lines.

A CV that parses correctly in one mode and garbles in the other has not passed. It has been lucky, once, on one machine.

Working rule behind every test on this site

Why you can check this yourself

The method uses two open-source packages and no proprietary service, so the measurements are reproducible by anyone with a PDF and a terminal. That is deliberate. If a number on this site is wrong, I would rather you find it: corrections are made in the article itself and the change date is updated, not quietly patched.

What I write about

• How applicant tracking software parses a CV, and which layout choices break it
• CV structure, section headings and file formats, measured rather than asserted
• Hiring practice differences between the UK, the DACH countries and the Netherlands, including the photo conventions that differ by market
• Job application automation, and where it genuinely does not help

Contact and corrections

Found an error, or measured something that contradicts a figure here? Say so on LinkedIn and I will either correct the article or publish why the original stands.