Back to projects

AI Tools & Applications

AI CV Screening — Explainable Match Scoring

Explainable CV-to-job match scoring that a recruiter can check line by line.

explainable score
0–100explainable score
labeled cases pass
33 / 34labeled cases pass
ranking checks pass
6 / 6ranking checks pass
model calls per match
≤ 2model calls per match

// at a glance

The problem and the idea

The problem

Screening CVs by hand is slow and inconsistent, and most AI screening tools return a number with no explanation. A black-box score can hide bias, reward keyword stuffing and is hard to defend to a candidate or a hiring manager.

The idea

Split the work. A language model only reads the CV and the job description and turns them into structured data. The score itself comes from plain, unit-tested Python rules, so every point traces back to a line in the reasons. Personal details are removed before any text leaves the service, and the result is decision support only: a person makes the call.

// who it's for

Who it's for

  • 01

    Recruiters and HR teams shortlisting many applicants for one role

  • 02

    Hiring managers who want an explained first pass before reading CVs

  • 03

    Candidates checking how well their CV fits a specific job

// how it works

How it works

CV · PDFdoc checkanonymizeLLM extractjob textPython scoringscore + reasons33 / 34 eval cases pass · no model in the score

1/5 Document check

Swipe sideways to see the whole diagram →

// screenshots

Screenshots

// key features

Key features

  • Evidence-weighted skills: a skill used in a job counts more than one only listed in a skills section
  • Experience computed from role dates, with overlapping roles merged
  • A warning when the CV copies text from the job posting
  • PDF upload or pasted text, with scanned PDFs detected
  • Per-IP rate limits, a daily model-call cap and a disk cache to control cost
  • Strict security headers; CV text is never written to disk or logged

// design decisions

Design decisions

  1. 01

    The LLM only extracts data; Python computes every number, so the score is reproducible and unit-tested.

  2. 02

    Evaluation labels were written before any model output existed, and no label was changed to make a case pass.

  3. 03

    Skill matching uses whole tokens, so Java never matches JavaScript and Go never matches MongoDB.

// built with

Built with

  • FastAPI
  • Python
  • OpenAI API
  • Pydantic
  • pypdf
  • Docker

Run it locally

$ uvicorn app.main:app --reload