An AI-powered Applicant Tracking System (ATS) resume analyzer — parses resumes, scores them against job descriptions using a hybrid of classic NLP/ML and LLM-based feedback, and gives candidates actionable suggestions to improve their match rate.
Most ATS-checker projects go one of two shallow routes: pure keyword matching (no semantic understanding) or a thin LLM wrapper (no real ML engineering). This project combines both deliberately:
- Lexical matching (TF-IDF/BM25) — literal keyword overlap, the way real legacy ATS systems work today.
- Semantic matching (sentence embeddings) — understands paraphrases and synonyms a keyword search would miss.
- Trained classifier — predicts shortlist likelihood from labeled data.
- LLM feedback layer — turns raw scores into specific, human-readable advice.
🚧 Actively in development. See commit history for daily progress.
See docs/architecture.md (coming soon) for the full system design and
reasoning behind each component.
- Backend: FastAPI, pdfplumber, python-docx, scikit-learn, sentence-transformers, spaCy
- Frontend: React
- Data: PostgreSQL + a vector store (Chroma/pgvector)
- Deployment: Docker, Render/Railway (backend), Vercel (frontend)
- Day 1: Text extraction (PDF/DOCX) + regex field extraction (email, phone) + rule-based section splitter
- Day 2: Lexical matching (TF-IDF + cosine similarity) and semantic matching (sentence embeddings) between resume and job description
- Day 3: Trained shortlist-prediction classifier (Logistic Regression, compared against Random Forest and Gradient Boosting) on engineered match features; combined lexical + semantic + classifier into a unified match report
- Day 4: LLM feedback module — turns match scores into specific, actionable suggestions, with an offline mock mode for development without an API key
- Day 5: Wired everything into a real FastAPI app (
/analyze,/healthendpoints), fixed an eager-model-loading bug that broke testability, added a test suite (health check + input validation run anywhere; full pipeline test requires network for the embedding model) - Day 6: React frontend (Vite) — resume/JD input form calling the backend
/analyzeendpoint, score breakdown display, suggestions list. CORS configured on the backend so the two can actually talk to each other locally. - Day 7: Auth + saved analysis history via Supabase (
/auth/signup,/auth/login,/history) — gracefully returns a clean 501 until real Supabase credentials are added, rather than crashing - Day 8: PDF report export (
/report/pdf) using reportlab — verified real, valid PDF output; snippet-only history storage decided deliberately for privacy - Day 9: Test coverage for the parsing and lexical scoring modules (previously untested) — email/phone extraction, section splitting, and TF-IDF relevance ordering
- Day 10: Dockerized both services — backend Dockerfile (trains the classifier fresh at build time since the model file is gitignored), multi-stage frontend Dockerfile (Node build → nginx serve), nginx configured to proxy
/apito the backend so the frontend code is identical between dev and production.docker-compose.ymlruns everything with one command.
docker compose up --buildFrontend: http://localhost:3000 · Backend API docs: http://localhost:8000/docs