GetJobCV

GetJobCV is a desktop application that attempts to score your resume against a given job description locally. It achieves this by implementing algorithms introduced as solutions in several white papers.

Usage

GetJobCV is currently Work-In-Progress. Do not expect the program to be functional. This project is an experiment to see how effective agentic tooling can be when working on small projects.

Roadmap

The roadmap follows the pipelines in the two referenced papers. Each step is tagged with its source: [P] Pimpalkar et al. (2023), Fig. 3 and Section III steps 1–6; [C] Chavan et al. (2024), Figs. 1–2 and Section 6.

1. Extract text from files [P step 1]

• Extract text from PDF
• Extract hyperlinks from PDF annotations
• Extract text from DOCX and RTF ("pdf, documents, images, and rich text formats")
• OCR for image-only or scanned resumes [P III.A]
• Use PDF layout (columns, tables, headers/footers) instead of plain reading order [P IV.4, IV.5, IV.7]

2. Preprocessing [P step 2, C Fig. 2]

• Lowercase, strip symbols and punctuation, tokenize, drop NLTK stopwords
• Keep tech tokens intact (C++, C#, .NET, Node.js)
• Stemming / lemmatization ("Data Prepare → Remove word stemming") [C Fig. 2]
• Section segmentation: split the resume into Education, Experience, Skills, Projects, etc. [P IV.6]
• Use sections in scoring: a matched skill counts fully when used (Experience, Projects, ...), 0.75 when mentioned (Education, Summary, ...), 0.5 when only in the Skills list

3. Information extraction: unstructured → structured [P III.A, C Section 6]

• Named Entity Recognition: people, organizations, locations (WikiNER)
• Skill gazetteer from O*NET plus a curated overlay, with demand tiers and the short names people write for O*NET's "... software" entries ("PLC", "HubSpot")
• Extract GitHub and LinkedIn socials
• Structured resume record: name, contact, education (degree, school, dates), job titles, companies, dates [P III.A]
• Years of experience, overall and per skill ("Skills and Experience showed significantly improved shortlisting") [C Section 6]
• Skill aliases → canonical names (JS → JavaScript, k8s → Kubernetes), taking the O*NET demand tier of any alias (alias<TAB>Canonical lines in Resources/skills.txt)
• Skill hierarchy: a specific skill covers a general one the job asks for (PostgreSQL → SQL, GitHub → Git → version control), at its own evidence and years (Specific > General lines in Resources/skills.txt)
• Parse the GitHub profile via the public API: repo languages and topics as a second skill source [C Fig. 1]
• Parse a LinkedIn profile from a user-supplied export (LinkedIn has no public profile API) [C Fig. 1]
• Merge CV, GitHub, and LinkedIn skill sets into one candidate skill set with its source for each skill [C Fig. 1]

4. Feature extraction and label encoding [P step 3]

• Count (bag of words)
• TF-IDF (IDF off for two-document comparison; on once a resume corpus exists)
• Cosine similarity
• Semantic features: word embeddings (GloVe/word2vec) for synonym-aware matching
• Doc2vec document vectors
• Label encoding of categorical fields (degree level, job category) for the classifiers

5. Resume classification [P step 4, C Section 6]

• Get a labeled resume dataset (resume text → job category)
• KNN job-role classifier on skill set, proficiency, and experience [C Section 6]
• Compare against Naïve Bayes, Logistic Regression, SVM, Random Forest, and Decision Tree [P step 4, Table 1]
• Classify the job description into the same categories and flag role mismatches

6. Ranking and shortlisting [P step 5, C Figs. 1–2]

• Weighted skill coverage (matched, missing, and extra skills by O*NET demand tier)
• Years-of-experience requirements from the job description ("3+ years of Python", "5+ years of experience"); matched skills lose credit when the resume shows fewer years
• Use the overall years requirement in the score: counts as one more requirement in skill coverage (weighted like a hot skill), credited by years shown / years required
• Split the job description into minimum and preferred qualifications: nice-to-have skills ("a plus", "Preferred Qualifications") count half (QualificationSplitter)
• Alternatives: a list where any one skill counts ("one of the following", "such as", "Python or Java", an inline topic list) is one requirement (AlternativeGroups)
• Staged ranking: score minimum credentials, then preferred credentials, then interview criteria, each stage with vectorization and cosine similarity [P step 5]
• Overall score: weighted blend of cosine match and skill coverage, with a verdict that's held back when the job has too few recognized skills (ScoreCombiner)
• Combine all stage scores and skill coverage into one final score
• Shortlist threshold: "similarities matched → CVs shortlisted / not shortlisted" [C Fig. 2]
• Run multiple resumes against one job and rank them [C Fig. 1 "Ranking Algorithm → Short-listed CVs"]
• Local candidate store so skill sets and scores are kept between runs [C Fig. 1 "Candidate DB"]
• Optional: TextRank-style resume summary (sentences → vectors → similarity matrix → graph → ranked sentences) [P Fig. 4]

7. Performance evaluation [P step 6]

• Hand-labeled test set of resume/job pairs (fit or no fit, expected skills), run through the whole pipeline (tests/GetJobCV.Tests/Evaluation)
• Precision, recall, and F1 for skill extraction; within-job concordance, AUC, and NDCG for fit ranking, with floors that fail on a regression
• Opt-in benchmark on the public resume/job fit dataset (not bundled: no license)
• Grow the gold set to 10 jobs and 21 resumes: DevOps, accounting, iOS, sales, electrical, and teaching next to tech and nursing
• Keep growing it with fields the gazetteer hasn't been tuned on
• Accuracy for classification and shortlisting, once those exist
• Unit tests for each module

8. Application

• Background processing and error handling
• Redo UI: start, progress, and report screens instead of a debug text dump
• Export the report as a PDF (printed with WebView2, the Edge engine built into Windows)
• Job-posting scraping as a job-description source ("web scraping") [C Fig. 2]

Out of scope for a local tool: the hosted-ATS features in [C] (interview scheduling, email automation, assessments, cloud deployment).

References

Chavan, Prasad & Chandurkar, Yash & Tidake, Ankita & Lavankar, Gaurav & Gaikwad, Suhani & Chavan, Rohit. (2024). Enhancing recruitment efficiency: An advanced Applicant Tracking System (ATS). Industrial Management Advances. 2. 6373. 10.59429/ima.v2i1.6373.

Pimpalkar, Amit & Lalwani, Aastha & Chaudhari, Roshan & Inshall, Mohd & Dalwani, Mahak & Saluja, Tarandeep. (2023). Job Applications Selection and Identification: Study of Resumes with Natural Language Processing and Machine Learning. 1-5. 10.1109/SCEECS57921.2023.10063010.

Copyright

The skill gazetteer is derived from the O*NET 30.3 Database by USDOL/ETA, used under CC BY 4.0. O*NET is a trademark of USDOL/ETA.