ResuMatch — Transparent, Fair Candidate Ranking (Not an ATS)

ResuMatch explores how AI can help rank resumes without skipping anyone.
Instead of auto-rejects, every applicant is parsed, scored, ranked — and explained.

Not an ATS: No silent filters. Everyone gets a score with visible rationale.

👀 At a glance (for reviewers)

• What it is: A lightweight Flask app that parses resumes + a JD, then ranks candidates.
• How it scores: A multi-factor model combining:
  • Semantic similarity (Sentence Transformers embeddings)
  • TF-IDF coverage (skills/keywords present)
  • Optional heuristics (e.g., experience indicators)
• Why it’s different: Transparency. Each ranking includes why it ranked that way.

⚖️ How ResuMatch Differs From an ATS

Most ATS (Applicant Tracking Systems):

• Auto-reject candidates if certain keywords are missing.
• Hide applicants from recruiters if they don’t match a predefined filter.
• Often act like a black box — candidates never know why they were skipped.

ResuMatch, by contrast:

• Never auto-rejects anyone — every resume is parsed and scored.
• Every candidate is ranked, not hidden.
• Transparency first: each score comes with an explanation breakdown (semantic similarity, TF-IDF coverage, heuristic signals).
• Built for fairness and insight, not filtering.

The goal is to augment recruiters, not replace human judgment.

🔎 How it works (high level)

1. Parse PDF/DOCX resumes and a job description.
2. Build vector embeddings for semantic comparison.
3. Compute TF-IDF coverage over required terms/skills.
4. Combine into a multi-factor score (tunable weights).
5. Rank candidates and output explanation notes.

🧰 Tech Stack

• Python · Flask
• Sentence Transformers (embeddings)
• scikit-learn (TF-IDF, metrics)
• PyPDF2 / python-docx (parsing)

📂 Key files

ResuMatch/ ├─ main.py # Flask entry + endpoints ├─ requirements.txt # Dependencies (CPU-friendly, numpy<2) ├─ utils/ # (If present) parsing + scoring helpers ├─ templates/, static/ # (If present) minimal UI └─ README.md

🧪 Example scoring snippet (core idea)

# Pseudocode for the core scoring
sem = cosine_sim(emb(job_desc), emb(resume_text))
tfidf_score = tfidf_coverage(resume_text, required_terms)
heur = simple_heuristics(resume_text)  # e.g., mentions of tools/years

score = 0.6 * sem + 0.35 * tfidf_score + 0.05 * heur
explanation = {
  "semantic_similarity": round(sem, 3),
  "tfidf_coverage": round(tfidf_score, 3),
  "signals": heur.details
}