AI-Powered Resume Screening and Job Recommendation System

Design a system that extracts useful information from diverse resumes and ranks job opportunities by relevance while limiting extraction errors, unfair bias, privacy risks, and misleading recommendations. 1.2 Objectives

• Extract skills, education, experience, and relevant terms from resumes.
• Represent job descriptions and candidate profiles in a comparable form.
• Generate explainable job-match scores and ranked recommendations.
• Evaluate extraction and ranking quality using documented test data.
• Include privacy, fairness, correction, and human-review safeguards. 1.3 Scope This paper describes an academic prototype for resume parsing and job discovery. It does not recommend rejecting applicants automatically, infer protected characteristics, or replace interviews, validated assessments, or recruiter judgment.

Information retrieval provides foundational methods for ranking documents by relevance. TF-IDF and cosine similarity are common baselines for matching resume text with job descriptions because they are simple and interpretable. They can underweight context and synonyms. Semantic embeddings may represent related phrases more closely, but must be evaluated for quality and bias. Resume parsing is sensitive to tables, columns, scanned pages, and inconsistent section headings. Job recommendation is affected by vacancy freshness, incomplete descriptions, location preferences, and changing career goals. Research on algorithmic decision-making also warns that historical hiring data can encode past inequalities; excluding protected attributes does not necessarily remove proxy effects.

A useful prototype should not focus only on ranking accuracy. It should combine parsing with explainable matching, candidate feedback, privacy-by-design, bias checks, and human oversight.

The candidate workflow allows a user to upload a resume, review extracted profile fields, set preferences, and receive ranked vacancies with explanations. The recruiter workflow allows an authorised user to create a vacancy and review matching profiles as one input to structured human assessment.

Figure 1. Conceptual system workflow: Resume Upload + Job Listings → Document Parsing → Text Cleaning & Skill Extraction → Candidate / Job Profiles → Matching and Ranking Engine → Explainable Recommendations → Human Review

A prototype may use Python and FastAPI for APIs, PostgreSQL for structured data, private object storage for uploaded documents, and React for the user interface. File processing should use size limits and isolated execution. Resumes must not be exposed through public URLs. Candidates should access their own profiles, and recruiters should only access applications they are authorised to review. 4. Methodology

Evaluation data can be assembled from consented, anonymised sample resumes and job descriptions whose usage terms permit analysis. Synthetic examples can test edge cases but should not be presented as representative real-world evidence. Remove duplicates, document sources and dates, and separate evaluation data by candidate or vacancy group to avoid nearduplicate leakage.

Validate file type, size, and extractable text before processing.

• Extract PDF/DOCX text and use OCR only for scanned documents.
• Normalise whitespace and section headings while retaining source text for correction.
• Identify skills with dictionaries, phrase matching, or a trained entity-extraction model.
• Mark uncertain fields for candidate verification.

A baseline can represent resume and vacancy text with TF-IDF word and phrase features and calculate cosine similarity. A semantic model may produce sentence embeddings. Structured features such as required-skill coverage, experience range, location, and employment type can be scored separately. Protected characteristics should not be used to rank candidates.

A transparent score can combine skill overlap, text relevance, experience fit, and explicit preferences: MatchScore = w1 × SkillCoverage + w2 × TextSimilarity + w3 × ExperienceFit + w4 × PreferenceFit. Weights should be selected using validation data and documented. The score is a relevance estimate, not a probability of job performance. Missing information should be shown as unknown rather than automatically treated as negative.

1. Receive and validate resume. 2. Extract text and candidate-approved fields. 3. Retrieve active jobs matching basic preferences. 4. Compute similarity and structured scores. 5. Rank vacancies and generate explanations. 6. Display results for review and correction. 7. Record only necessary feedback and audit events. 8. Evaluate ranking quality and fairness periodically.

Collect only necessary data and obtain informed consent.

• Encrypt data in transit and at rest; restrict resume access by role.
• Allow correction of extracted details and deletion where applicable.
• Set retention periods for resumes, profiles, and feedback.
• Do not expose personal details or credentials in logs.
• Keep human review for consequential hiring decisions.

7.1 Advantages

• Reduces repetitive job searching and organises large job catalogues. • Supports skill discovery across different resume wording. • Explains matching factors and allows profile correction. • Uses modular components that can be tested independently. 7.2 Limitations • Scanned or poorly formatted resumes can cause extraction errors. • Similarity does not measure true ability or future job performance. • Incomplete or outdated job descriptions produce weak recommendations. • Models may contain bias or fail on unfamiliar terms. • Privacy, accessibility, and fairness require continuous review.

Recruitment data is sensitive and can affect employment opportunities. Provide clear notice about data use and retention. Candidates should inspect and correct extracted information. Recruiters should not treat ranking as an objective verdict. Protected attributes and inferred personal characteristics should not determine suitability. Real hiring use should be reviewed against applicable law and organisational policy.

Future work may include multilingual parsing, improved OCR, semantic retrieval, candidate-controlled preferences, vacancy freshness checks, duplicate-job detection, and accessible interfaces. Further study may evaluate counterfactual fairness, ranking calibration, privacy-preserving analytics, and user experience. Extensions should be validated with representative data and independent review. 10. Conclusion This paper proposed an AI-powered resume screening and job recommendation system combining resume parsing, skill extraction, text and semantic matching, ranked job discovery, and human review. It emphasises explainability, correction, privacy, fairness evaluation, and reproducible metrics. The design can support an academic prototype, but its effectiveness and suitability for recruitment must be demonstrated through careful testing and responsible governance.