Hirely — Resume Screener

Hirely is a portfolio-grade resume screening platform built with Streamlit and a hybrid NLP ranking engine. It parses resumes, extracts skills from a large skill taxonomy, computes semantic relevance to a job description, and ranks candidates with transparent scoring and persisted results in SQLite.

Project Overview

Hirely helps recruiters and hiring teams:

• Upload and parse PDF resumes.
• Store jobs, candidates, and ranking outputs in a local database.
• Rank candidates using hybrid scoring (semantic relevance + required skill coverage).
• Inspect candidate-level explanations for decision support.
• Evaluate ranking quality with benchmark metrics.

Features

• Multi-page Streamlit dashboard for end-to-end screening workflow, including a Ranker Evaluation page.
• Dynamic skill extraction using a curated JSON taxonomy (data/skills.json) with 236 real-world skills.
• Hybrid ranking engine:
  • Semantic similarity (TF-IDF + cosine similarity)
  • Skill match ratio against required skills
• Persistent storage via SQLite (hirely.db) with tables for jobs, candidates, results.
• Evaluation module with Precision@K, Recall@K, and MRR.
• Explainable outputs showing semantic score, skill match %, and missing skills.

System Architecture

User Upload Resume
        ↓
PDF Parser (PyMuPDF)
        ↓
Text Preprocessing
        ↓
Skill Extraction (skills.json taxonomy)
        ↓
Semantic Similarity Engine (TF-IDF + Cosine)
        ↓
Hybrid Ranking Engine
        ↓
SQLite Storage (jobs, candidates, results)
        ↓
Streamlit Dashboard

NLP Pipeline

1. PDF parsing (resume_parser.py) extracts text from uploaded resumes.
2. Preprocessing (ml_pipeline.py) normalizes and tokenizes text.
3. Skill extraction (skill_extractor.py) matches normalized terms from data/skills.json.
4. Semantic scoring computes cosine similarity between the job text and each candidate profile.
5. Hybrid ranking blends semantic similarity and skill coverage.

Scoring Algorithm

For each candidate:

• semantic_similarity ∈ [0, 1]
• skill_match_ratio ∈ [0, 1]

Final score:

final_score = (0.7 * semantic_similarity + 0.3 * skill_match_ratio) * 100

This avoids misleading scaling and ensures normalization to 0–100 only at the end.

Data Model & Persistence

SQLite tables

• jobs: title, description, required skills
• candidates: parsed resume text, cleaned text, extracted skills
• results: final scores + explainability fields

Database helper module: database.py.

Evaluation & Benchmark

evaluation.py includes a mini benchmark dataset with expected relevant candidates and computes:

• Precision@K
• Recall@K
• Mean Reciprocal Rank (MRR)

Current benchmark output (python evaluation.py, k=2):

precision@k	recall@k	mrr
0.75	1.00	1.00

Run:

python evaluation.py

Example Ranking Output

Rank	Candidate	Semantic %	Skill Match %	Final Match %	Missing Skills
1	Alice Johnson	84.3	100.0	88.99	None
2	Ben Torres	73.6	80.0	75.52	Hugging Face Transformers
3	Carla Smith	21.0	0.0	14.70	Python, SQL, NLP, MLOps

Screenshots

Add screenshots by running the app locally and capturing Dashboard, Ranking, and Insights pages.

Setup Instructions

pip install -r requirements.txt
streamlit run app.py

Project Structure

hirely/
├── app.py
├── ml_pipeline.py
├── resume_parser.py
├── database.py
├── evaluation.py
├── skill_extractor.py
├── utils.py
├── data/
│   └── skills.json
├── requirements.txt
└── README.md

Why this is portfolio-ready

• Clean separation of UI, NLP pipeline, and persistence.
• Explainable and reproducible ranking outputs.
• Benchmark-oriented evaluation for technical review.
• Lightweight and deployable architecture without heavyweight infra.