Resume Screening AI Agent

📖 Abstract

Resume Screening AI Agent is a production-ready Streamlit application that intelligently scores and shortlists resumes against job descriptions using local SentenceTransformers embeddings, Ollama-powered LLM reasoning, and a lightweight NumPy vector store. The system combines semantic similarity matching with skill and experience heuristics to provide comprehensive candidate evaluation with detailed strengths, weaknesses, and reasoning for each applicant. This project demonstrates advanced NLP techniques, AI agent architecture, and practical HR automation solutions.

---

🛠️ Tech Stack Layer Technology Framework Streamlit LLM Ollama (Gemma3:1b) Embeddings SentenceTransformers (SBERT) Vector Store NumPy + JSON Document Parsing pdfplumber Language Python 3.10+ Storage Local File System

📌 Project Summary Section Details Project Name Resume Screening AI Agent Developer Srinivas B N College RNSIT, Bengaluru Branch CSE-AIML (2026) Challenge Rooman 48-Hour AI Agent Challenge Stack Local Ollama + SBERT (Free) Purpose Automated resume screening and ranking

---

✅ Features

Core Functionality Upload JD + multiple resumes (PDF/TXT) pdfplumber-based parsing Whitespace normalization Local SBERT embeddings Cosine similarity matching Skill detection heuristics Experience evaluation

AI-Powered Analysis Ollama LLM integration Candidate strengths analysis Weakness identification Detailed reasoning per candidate Semantic similarity scoring Multi-layer ranking algorithm Context-aware evaluation

Output & Storage Downloadable shortlist CSV Expandable candidate cards Persistent NumPy vector store Modular architecture Comprehensive logging Error handling system Session-based storage

---

🎯 Current Outcome

This project currently contains:

• ✔ Production-ready Streamlit application
• ✔ Local Ollama + SBERT integration (no API costs)
• ✔ Multi-resume batch processing
• ✔ Intelligent ranking algorithm
• ✔ Detailed candidate insights
• ✔ CSV export functionality
• ✔ Modular and extensible codebase

---

🚀 Quick Start Guide

1️⃣ Clone Repo

git clone https://github.com/SRINIVASBN/Resume-Screening-AI-Agent.git
cd Resume-Screening-AI-Agent

2️⃣ Install Ollama

1. Download Ollama from https://ollama.com
2. Install and start Ollama:

ollama serve

3. Pull the Gemma3 model:

ollama pull gemma3:1b

3️⃣ Setup Python Environment

# Create virtual environment
python -m venv venv

# Activate (Windows)
.\venv\Scripts\Activate.ps1

# Activate (Mac/Linux)
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

4️⃣ Configure Environment (Optional)

# Windows PowerShell
$env:OLLAMA_URL = "http://127.0.0.1:11434/api/generate"
$env:OLLAMA_MODEL = "gemma3:1b"

# Mac/Linux
export OLLAMA_URL="http://127.0.0.1:11434/api/generate"
export OLLAMA_MODEL="gemma3:1b"

5️⃣ Run the Application

streamlit run app/main.py

---

🏗️ Architecture

flowchart LR
    subgraph UI[Streamlit UI]
        JD[JD Upload]
        RES[Resume Uploads]
        TABLE[Shortlist Table]
    end

    JD -->|save| FM[File Manager]
    RES -->|save| FM
    FM --> PARSER[Document Parser]
    PARSER --> TEXTS[Clean Text Corpus]
    TEXTS -->|JD| EMB[SBERT Embedding Service]
    TEXTS -->|Resumes| VECTOR[NumPy Vector Store]
    EMB --> VECTOR
    EMB --> MATCHER[Candidate Scorer]
    VECTOR --> MATCHER
    MATCHER --> LLM[LLM Client]
    LLM --> MATCHER
    MATCHER --> TABLE

Loading

---

📁 Project Structure

Resume-Screening-AI-Agent/
├── app/
│   ├── main.py              # Streamlit entrypoint
│   ├── embeddings/          # Embedding + vector-store management
│   ├── parsing/             # File parsing utilities
│   ├── prompts/             # Prompt templates
│   ├── ranking/             # Scoring + ranking logic
│   ├── utils/               # Logging, file IO, LLM client, helpers
│   └── storage/             # Local persistence
│       ├── uploads/         # Uploaded files
│       ├── chroma/          # Vector embeddings
│       └── app.log          # Application logs
├── requirements.txt         # Python dependencies
└── README.md               # Documentation

---

🔧 How It Works

Step 1: Document Upload & Parsing

• Users upload Job Description (JD) and multiple resumes
• pdfplumber extracts text with whitespace normalization
• Documents are cleaned and preprocessed

Step 2: Embedding Generation

• Local SentenceTransformers (SBERT) generates embeddings
• Embeddings stored in NumPy-based vector store
• Fast cosine similarity computation

Step 3: Candidate Scoring

• Multi-layer scoring algorithm:
  • Semantic similarity matching
  • Skill detection heuristics
  • Experience evaluation
• Weighted scoring system

Step 4: LLM Analysis

• Ollama generates detailed insights:
  • Candidate strengths
  • Areas of improvement
  • Reasoning for scores
• Context-aware evaluation

Step 5: Results & Export

• Ranked shortlist displayed in table
• Expandable cards with detailed insights
• Downloadable CSV for offline review

---

🎨 Key Components

Document Processing Parser: pdfplumber-based extraction Normalization: Whitespace cleaning Format Support: PDF, TXT Error Handling: Graceful fallbacks

AI & Embeddings Model: SentenceTransformers (SBERT) LLM: Ollama Gemma3:1b Similarity: Cosine similarity Storage: NumPy vector store

Scoring Algorithm Layer 1: Semantic similarity Layer 2: Skill matching Layer 3: Experience heuristics Output: Weighted composite score

---

☁️ Deployment (Streamlit Cloud)

1️⃣ Push to GitHub

git push origin main

2️⃣ Create Streamlit Cloud App

1. Go to share.streamlit.io
2. Connect your GitHub repository
3. Set main file path: app/main.py

3️⃣ Configure Secrets

Add to Streamlit Cloud secrets:

OLLAMA_URL = "https://your-ollama-host/api/generate"
OLLAMA_MODEL = "gemma3:1b"

4️⃣ Deploy

Click "Deploy" and your app will be live!

---

📊 Testing the Flow

1. Upload JD: Upload job description (PDF/TXT)
2. Upload Resumes: Add multiple candidate resumes
3. Run Screening: Click "Run Screening" button
4. Review Results: Examine shortlist table
5. Expand Cards: View detailed candidate insights
6. Download CSV: Export results for offline review

---

🔒 Logging & Storage

Application Logs

• Location: app/storage/app.log
• Rotation: 1 MB per file
• Format: Timestamped with log levels

File Storage

• Uploads: app/storage/uploads/
• Vectors: app/storage/chroma/
• Artifacts: Session-based persistence

---

⚠️ Limitations

• Ollama Dependency: Requires local Ollama instance running
• File Formats: Currently supports PDF and TXT only
• Layout Complexity: Complex PDF layouts may degrade extraction
• Skill Detection: Based on predefined dictionary
• Vector Store: Local persistence (not suitable for multi-user deployments)

---

🚀 Future Improvements

• File Format Support: Add DOCX, HTML with fallback parsers
• Resume Parsing: Extract names, contact info via NLP tagging
• Skill Taxonomy: User-customizable keyword library
• Feedback Loop: Store reviewer decisions for model tuning
• Enterprise Features: Dockerfile, CI/CD workflows
• Cloud Storage: Migrate to managed vector databases
• Multi-language Support: International resume screening
• Advanced Analytics: Dashboard with hiring metrics

---

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

---

👨‍💻 Author

SRINIVAS BN
CSE-AIML (2026)
RNSIT, Bengaluru

---