AI Resume Screening System

Parse, Match and Rank Job Applications

Design a system that turns uploaded resumes into structured profiles, ranks applicants against a job with reasons a recruiter can check, and proves with numbers that the ranking does not disadvantage any group.

Overview

A popular job can draw thousands of applications. Recruiters cannot read them all with equal care, so companies use software that reads each resume, compares it with the job, and puts the most relevant applicants first. This problem is that software: an asynchronous pipeline that parses resumes, a scorer that ranks them, and the controls that keep the ranking fair and explainable.

The technical parts are familiar: file parsing, text extraction with a language model, embeddings, a ranked list per job. What makes the problem different is the risk. A screening model learns from what recruiters did before, and if they favored one group, the model learns that too. Reuters reported in 2018 that Amazon dropped an experimental recruiting engine after it learned to penalize resumes containing the word 'women's', because it had been trained on ten years of mostly male applicants. Laws now reach these tools directly: New York City requires a yearly independent bias audit before an employer uses one, and the EU AI Act lists AI that filters job applications as high-risk.

Questions to ask first

Does the system reject applicants on its own? — No. It ranks and explains; a recruiter decides
What volume? — 2 million applications a day across 200,000 open jobs
How fast must a new application be scored? — Within 5 minutes of upload
Which file types? — PDF, Word and plain text; scanned PDFs need OCR
Where is it used? — US and EU employers, so NYC Local Law 144 and the EU AI Act apply
Do we have demographic data? — Only voluntary self-identification, stored apart from the profile and never used to score

What you will be able to do

Build the parse and score pipeline and say why it is asynchronous.
Separate must-have requirements from ranking signals, and explain every score with evidence from the resume.
Name the proxy features that leak protected traits and say what removing them does and does not fix.
Compute adverse impact with the four-fifths rule in SQL and say what to do when a group falls below it.
Explain why recruiter decisions are biased labels and how to retrain safely.

Key concepts

Resume parsing
Skill normalization
Must-have vs nice-to-have
Semantic matching
Explainable scores
Proxy features
Adverse impact (four-fifths rule)
Human in the loop
Audit trail

Functional requirements

Accept resumes in PDF, Word and text, and parse them into a structured profile (roles, dates, skills, education)
Score each application against its job: must-have checks plus a ranking score
Show a ranked list per job with the evidence behind each score
Let recruiters record advance or reject decisions with a reason code
Re-score all applicants when the job description changes
Produce adverse-impact reports by group and an audit log of every score and decision

Non-functional requirements

A new application is scored within 5 minutes at p95 (a target we chose)
The ranked list for a job loads in under 1 second for 5,000 applicants (a target we chose)
Every score is reproducible: the model version and inputs are stored with it
Demographic data is never an input to scoring and is readable only by the audit job and compliance staff
No application is lost: uploads are durable before the API answers

Interview questions

How do you keep the system from discriminating?
Walk me through the four-fifths rule with numbers.
Why is scoring asynchronous instead of done in the upload request?
How do you explain a candidate's rank to a recruiter?
Can you just train a model on who got hired?
What changes for an employer hiring in New York City or the EU?
How would you stop keyword stuffing?