Resume Screening using NLP

A single popular job posting can attract hundreds of resumes, and manual screening does not scale: recruiters skim each resume for seconds, strong candidates get missed, and two screeners can rank the same pile differently. Keyword-only applicant tracking systems help, but they are brittle — a candidate who writes "RESTful services" instead of "REST API" can be filtered out despite being qualified. This project builds a more principled screener: resumes and the job description are normalized through an NLP preprocessing pipeline (tokenization, stopword removal, stemming), required skills are extracted with a domain lexicon, and each resume is scored by combining TF-IDF cosine similarity with skill-overlap against the JD. The result is a ranked shortlist with per-candidate explanations — matched skills, missing skills, and similarity scores — so the screening decision stays with the human, but the grunt work is automated and auditable.

1. The recruiter pastes a job description; the lexicon extractor identifies the required skill set.
2. Each resume is preprocessed: lowercasing, tokenization, stopword removal and stemming.
3. TF-IDF vectors are built over the JD plus all resumes, so term weights reflect the actual applicant pool.
4. Cosine similarity between each resume vector and the JD vector gives the text-similarity component.
5. Skill overlap (matched required skills ÷ total required skills) gives the skill component; the two combine into the final score.
6. Candidates are ranked, labeled with verdicts, and each row expands into the inspector view with highlighted evidence.

Tech stack:

• Python 3.10, scikit-learn (TfidfVectorizer, cosine_similarity)
• Regex + domain skill lexicon (44 terms: languages, frameworks, cloud, DevOps)
• NLTK-style preprocessing (tokenize, stopwords, Porter stemming)
• Flask demo app (JD input, ranking dashboard, resume inspector)
• pandas (shortlist tables and batch exports)
• Matplotlib (TF-IDF weight charts, similarity heatmaps)
• Jupyter notebook (buyer-run pipeline walkthrough and evaluation)

• [Job description skill extraction] Required skills are pulled from any pasted JD with a 44-term domain lexicon covering languages, frameworks, databases, cloud and DevOps tooling, so the "required set" adapts to each posting.
• [TF-IDF resume ranking] Resumes are vectorized with TF-IDF and ranked by cosine similarity to the JD, weighted 35% alongside 65% skill-overlap for a transparent composite score.
• [Skill-gap analysis] Every candidate gets a matched-vs-missing skill breakdown, showing exactly which required skills would close the gap.
• [Resume inspector] Click any ranked candidate to see the parsed resume with matched keywords highlighted and the extracted profile (title, experience, contact, skills).
• [Similarity heatmap] A resume-to-resume cosine similarity matrix reveals which candidates have similar backgrounds — useful for diversifying a shortlist.
• [Verdict thresholds] Configurable cutoffs label candidates Shortlist, Review or Reject, with the thresholds documented and adjustable in code.
• [Batch screening] Score a whole folder of resumes in one run and export the ranked shortlist as a report table.
• [Evaluation notebook] A Jupyter notebook walks through preprocessing, TF-IDF weighting and ranking on your own resume set, with precision-at-k computed on your labels.

• Complete source code (preprocessing, TF-IDF pipeline, skill extractor, scoring, Flask demo)
• Jupyter notebook (pipeline walkthrough + precision-at-k evaluation procedure on your resume set)
• Project report PDF (background, NLP design, scoring formula, evaluation, bias discussion)
• PPT presentation for final review
• Viva Q&A preparation document (TF-IDF, cosine similarity, stemming, precision-at-k, limitations)
• Setup guide (environment, preparing your resume set, running batch screening)

• The system ranks resumes against the JD you provide — it cannot judge qualities absent from text (culture fit, communication, real ability), and the report states this boundary explicitly.
• Skill extraction depends on lexicon coverage: synonyms or brand-new tools outside the 44-term list are missed unless you extend the lexicon, which the guide shows how to do.
• TF-IDF is lexical, not semantic — "RESTful services" and "REST API" score as different terms; the report discusses this as the motivation for embedding-based future scope.
• PDF resumes without a text layer (scanned images) need OCR first, which is outside the base pipeline.
• Any screening aid can amplify historical bias if trained or tuned on past hiring decisions — the report includes a bias-and-fairness discussion, and the tool is positioned as a screener's assistant, never an autonomous hiring decision-maker.