Skip to content

README.md

Latest commit

 

History

History
139 lines (101 loc) · 4.46 KB

README.md

File metadata and controls

139 lines (101 loc) · 4.46 KB

SkillScope

Most job advice tells you to "build in-demand skills." Almost none of it tells you which ones, for which roles, at which companies, right now.

Course syllabi lag. Generic skill lists are written by people who aren't hiring. SkillScope skips both and reads the job postings directly.

🔗 Live Demo: aashshahh.github.io/skillscope

The Problem

You upload your resume to a job application. It goes into a black box. You hear nothing.

Part of what's happening in that black box is a skill match your resume against what the job description actually requires. The problem is you never see that comparison. You don't know which skills tipped you out of the running, which ones you already have that they care about, or what to go learn before the next application.

SkillScope makes that comparison visible.

Paste your resume and a job description. Within seconds you get:

  • every skill the JD requires, extracted and normalized
  • which of those you already have (exact matches)
  • which you're close on similar concepts, different phrasing (semantic matches via Sentence-BERT cosine similarity)
  • what's genuinely missing, ranked by how central it is to the role

Not keyword overlap. Actual semantic distance.

How It Works

Skill Extraction

Job descriptions and resumes don't use consistent language. One posting says "proficiency in Postgres," another says "strong SQL background," a third lists "PostgreSQL" under requirements. A keyword matcher treats these as three different things. SkillScope normalizes them to the same canonical skill using a curated taxonomy mapped to the ESCO ontology, then runs spaCy's PhraseMatcher across 150+ skills across ten categories.

Semantic Matching

Exact match only gets you so far. If your resume says PyTorch and the JD says TensorFlow, that's not a zero — it's a partial signal. SkillScope embeds every extracted skill using Sentence-BERT (all-MiniLM-L6-v2), computes a cosine similarity matrix between your skills and the JD's skills, and flags anything above a 0.65 threshold as a soft match. You see the score, the pairing, and can judge it yourself.

Gap Score

The final output is a 0–100 match score: exact matches count fully, soft matches count proportionally by their similarity score. It's not a vibe. It's a number you can act on.

Tech Stack

  • NLP: spaCy, Sentence-Transformers, NLTK
  • Skill taxonomy: ESCO v1.1 + curated skills.json (150+ skills, 10 categories)
  • Backend: FastAPI, Uvicorn, Pydantic
  • Frontend: React 18, Vite, no UI framework — custom CSS
  • Evaluation: SkillSpan dataset (HuggingFace) — precision, recall, F1
  • Data: RemoteOK API (live job scraping), synthetic resumes
  • CI/CD: GitHub Actions (test + deploy)
  • Deployment: GitHub Pages (frontend), Render (API)

Getting Started

# Clone
git clone https://github.com/aashshahh/skillscope.git
cd skillscope

# Backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python -m spacy download en_core_web_sm

# Test the pipeline directly
python scripts/run_pipeline.py \
  --resume data/raw/sample_resume.txt \
  --jd data/raw/sample_jd.txt

# Start the API
uvicorn src.api.main:app --reload

# Frontend (separate terminal)
cd frontend && npm install && npm run dev

Open http://localhost:5173. The API runs at http://localhost:8000/docs.

Project Structure

skillscope/
├── src/
│   ├── ingest/       # resume parser (PDF, DOCX, TXT), job scraper
│   ├── nlp/          # skill extractor, SBERT embedder, normalizer
│   ├── models/       # gap analyzer, evaluator
│   └── api/          # FastAPI routes + schemas
├── frontend/         # React + Vite
├── scripts/          # CLI pipeline runner, evaluation script
├── config/
│   ├── settings.yaml
│   └── skills.json   # skill taxonomy
├── data/raw/         # sample resume + JD for testing
└── tests/            # pytest unit tests

Evaluation

Run against the SkillSpan benchmark dataset (HuggingFace):

python scripts/evaluate.py --n 200 --output results/eval_report.json
Model Precision Recall F1
spaCy PhraseMatcher + ESCO 0.87 0.84 0.85
BERT baseline 0.82 0.79 0.80

Aash Shah · LinkedIn · aashshah.04@gmail.com