How tanyaignatenko/resume-parser works
Project Type & Purpose
• This is a full-stack AI-powered Resume Parser.
• Frontend: Next.js/React web app for upload & display.
• Backend: FastAPI service hosting a custom spaCy NER model.
• Training: spaCy training pipeline driven by Python scripts and an annotated resume dataset.
• Deployment: Frontend + Backend on Vercel (backend configured via vercel.json).
File-Structure Snapshot & Architectural Clues
• backend/
• main.py, api/index.py → FastAPI entrypoints
• model/… → spaCy model artifacts (cfg, vocab, tokenizer, ner weights)
• requirements.txt, vercel.json → dependencies & deployment config
• frontend/
• app/** → Next.js pages (upload page, results page)
• components/** → React UI components & custom UI library
• lib/api.ts → client utilities to call the backend
• next.config.mjs, package.json → build & framework config
• training/
• annotated/*.json → train/dev/test splits
• *.py scripts → data preprocessing, pre-labeling, spaCy training
• No external database—model artifacts and data live in-repo; uses pdfplumber & docx2txt for file parsing.
Key Architecture Elements
a. Frontend (Next.js + React)
• UI pages: Upload form, Results view
• API client: POST “/parse” endpoint
b. Backend (FastAPI + Uvicorn)
• API Layer: parse endpoint defined in api/index.py
• Processing Pipeline:
1. File ingestion (PDF/DOCX/TXT)
2. Text extraction via pdfplumber/docx2txt
3. Text cleaning & normalization
4. Tokenization & spaCy NER inference (load from backend/model)
5. Schema mapping (Skills, Experience, Education, Languages) → JSON
c. Model Storage
• spaCy model directory containing cfg, tokenizer, vocab, ner weights
d. Training Pipeline (Offline)
• Annotation: Label Studio → annotated/*.json
• Preprocessing & pre-label scripts (ls_to_spacy.py, make_prelabels.py)
• Training: train_ner.py using spaCy’s CLI or API
• Evaluation: show_mismatches.py, reporting F1/Precision/Recall