ScreenAI — Automated Recruitment Screening System

An AI-powered CV screening platform that automates early-stage candidate selection using Blind Screening (NER-based anonymization), Competency Evaluation (RAG pipeline), and Explainable AI (evidence-based score justifications). Built for fair, transparent, and efficient recruitment — designed as a local single-user system.

---

Tech Stack

Layer	Technology
Backend	FastAPI + SQLAlchemy + SQLite
PDF Parsing	PyMuPDF
NER (Anonymization)	IndoBERT (ageng-anugrah/indobert-large-p2-finetuned-ner)
RAG & Orchestration	LangChain + ChromaDB
LLM Inference	DeepSeek V3 (OpenAI-compatible API)
Frontend	React + Vite + Tailwind CSS + shadcn/ui

---

Prerequisites

• Python 3.10+
• Node.js 18+
• DeepSeek API key (get one at platform.deepseek.com)
• ~4 GB disk space (for IndoBERT model cache)

---

Setup

1. Clone the repository

git clone https://github.com/istgrudd/screenai.git
cd screenai

2. Backend

# Create and activate virtual environment
python -m venv venv
source venv/bin/activate       # Linux/macOS
venv\Scripts\activate          # Windows

# Install dependencies
pip install -r requirements.txt

# Configure environment
cp .env.example .env
# Open .env and fill in your DEEPSEEK_API_KEY

Run the backend:

uvicorn backend.main:app --reload --port 8000

API will be available at http://localhost:8000. Health check: GET /api/health.

3. Frontend

cd frontend
npm install
npm run dev

UI will be available at http://localhost:5173.

4. Seed default rubric

python -m scripts.seed_rubric

This populates the database with a sample rubric for testing.

---

Project Structure

screenai/
├── backend/
│   ├── main.py              # FastAPI entry point
│   ├── models/              # SQLAlchemy ORM models
│   ├── routers/             # API route handlers
│   ├── services/            # Core pipeline logic
│   │   ├── extractor.py     # PDF → raw text (PyMuPDF)
│   │   ├── anonymizer.py    # NER-based identity masking
│   │   ├── rag_pipeline.py  # LangChain RAG orchestration
│   │   └── scoring.py       # Weighted score aggregation
│   └── utils/               # Helpers (PDF, NER, LLM client)
├── frontend/
│   └── src/
│       ├── pages/           # Upload, Dashboard, Detail, Rubric
│       └── components/      # UI components
├── scripts/
│   └── seed_rubric.py       # Seed default rubric
├── data/                    # Local data (gitignored)
├── requirements.txt
├── .env.example
└── CLAUDE.md                # Execution plan

---

Features

• PDF Upload — Multi-file upload for CVs and language certificates
• Text Extraction & Normalization — Section-aware parsing (education, experience, skills, certifications)
• Blind Screening — Automatic anonymization of names, contacts, institutions via IndoBERT NER + regex fallback
• Rubric Configuration — Define competency dimensions, weights, and indicators per job position
• RAG-based Scoring — Embed rubric → retrieve relevant CV chunks → generate structured scores via DeepSeek V3
• Explainable AI — Every score includes specific CV evidence and a justification
• Profile Summary — Narrative candidate overview generated by LLM
• Candidate Ranking — Dashboard sorted by weighted composite score
• Score Override — Recruiters can manually adjust dimension scores
• Batch Processing — Evaluate all candidates for a rubric in one action

---

Team

Kelompok 26 — Capstone Design, Telkom University 2026