I Built My Own Tiny ATS to Stress-Test the "AI Scoring" Resume Advice — Here's What Held Up and What Didn't

Follow-up to Your Job Application Is Being Scored by AI and I Reverse-Engineered 5 ATS Systems . Both posts are based on testing real resumes against real ATSes. But there's only so much you can learn from the outside. So I built a tiny ATS scorer myself — semantic match + keyword density + section parser — and ran 50 resumes through it. Here's what the conventional advice gets right, what it gets wrong, and what nobody is saying.

What I built (300 lines, weekend project)

Section parser — splits a resume PDF into Header / Summary / Experience / Skills / Education using regex + heuristics. Same approach Greenhouse and Lever use under the hood (their parsers are public-ish — search GitHub for resume parser and you find the patterns).

Keyword scorer — pulls a job description, extracts top 30 phrases, checks each against the parsed resume.…