🧠 LinkedIn Apply Bot with Offline AI Integration (Ollama)

Project Overview

This project automates the job application process on LinkedIn Easy Apply and extends support to external platforms such as Greenhouse, Ashby, and more. It features AI-powered resume tailoring and intelligent question answering via offline LLMs (Ollama or any LLM provider through LiteLLM), delivering a personalized, end-to-end automation experience for job seekers.

---

🚦 Project Status

✅ Stable Features

• Automated login and job application for LinkedIn Easy Apply
• AI-driven question answering for application forms (text, numeric, multiple-choice)
• RAG-powered resume context optimization with semantic chunking and vector retrieval
• Resume tailoring and skill replacement to optimize ATS scores
• Support for both PDF and DOCX resumes
• Offline AI integration (Ollama + phi4-mini) for data privacy and speed
• Flexible backend—integrate any LLM via LiteLLM
• Modular codebase designed for extensibility and additional platforms
• Semantic search using lightweight embedding models for better context relevance

🧪 Experimental Features

• Support for external applications including Greenhouse and Ashby
• Resume tailoring and regeneration (DOCX to PDF)

🚧 In Development / On the Way

• Improved resume tailoring using RAG with semantic chunking and vector retrieval ✅ COMPLETED
• Lightweight LLM-powered rewriting of resume sections based on job descriptions
• Resume context compression for small models to reduce hallucination ✅ COMPLETED
• Confidence scoring and APPLY/SKIP justification for job fit evaluation
• Pluggable model backend (phi4-mini, Mistral, TinyLlama, etc.)
• Logging outcomes for feedback loop and model fine-tuning

🧠 RAG (Retrieval-Augmented Generation) Features

Smart Context Building

The bot now uses advanced RAG techniques to optimize context for small LLMs:

• Semantic Chunking: Resume content is intelligently split into meaningful chunks by sections and context
• Vector Embeddings: Uses all-MiniLM-L6-v2 (90MB) for fast, accurate semantic search
• FAISS Index: Lightning-fast similarity search for relevant resume sections
• Query-Aware Retrieval: Finds most relevant resume content based on job description and questions
• Context Caching: Avoids recomputation for better performance

Benefits for Small LLMs

• Reduced Token Usage: Only sends relevant context instead of entire resume
• Better Accuracy: Focused information reduces hallucination
• Faster Response: Less context to process means quicker generation
• Cost Effective: Fewer tokens = lower API costs
• Scalable: Works efficiently with resumes of any length

Technical Implementation

# Semantic search for relevant resume sections
relevant_chunks = self._semantic_search(query + job_description, top_k=8)

# Context optimization for small models
context = self._build_context_rag(
    query=question_text, 
    job_description=jd, 
    max_tokens=1500  # Optimized for small LLMs
)

🧩 Technologies Used

• Python (Selenium, PDF, DOCX)
• Ollama for offline LLM chat
• phi4-mini model
• Sentence Transformers for semantic embeddings
• FAISS for vector similarity search
• NumPy for efficient vector operations
• PyAutoGUI (to prevent system sleep)
• Regex, JSON, CSV, and automation utilities

📦 Getting Started

Quick Start

1. Clone the repository

   git clone https://github.com/pratikjadhav2726/LinkedInEasyApplyBot.git
   cd LinkedInEasyApplyBot

2. Install dependencies (requires uv)

   uv sync --dev

3. Configure the bot

   cp examples/config.yaml.example config.yaml
   # Edit config.yaml with your LinkedIn credentials and preferences

4. Run the bot

   uv run python src/main.py

For detailed setup instructions, see docs/SETUP.md.

🚀 Installation & Usage with UV (Recommended)

This project uses UV for fast and reliable dependency management.

Prerequisites

• Python 3.9 or higher
• UV package manager (install UV)

Quick Setup

# Clone the repository
git clone https://github.com/pratikjadhav2726/LinkedInEasyApplyBot.git
cd LinkedInEasyApplyBot

# Install all dependencies (runtime + dev/test)
uv sync --dev

# Run the bot
uv run python main.py

RAG-Specific Dependencies

# Core RAG dependencies
uv add sentence-transformers  # For embeddings
uv add faiss-cpu             # For vector search
uv add numpy                 # For vector operations

Common UV Commands

# Install dependencies
uv sync

# Install dependencies including dev/test groups
uv sync --dev

# Add a new dependency
uv add package-name

# Add a development dependency
uv add --dev package-name

# Remove a dependency
uv remove package-name

# Run the project
uv run python main.py

# Run with script entry point
uv run linkedin-bot

# Export pinned requirements (compatibility with legacy tools)
uv export --frozen --no-dev > requirements.txt

# Update dependencies
uv lock --upgrade

Key Files

• pyproject.toml - Project configuration and dependencies
• uv.lock - Lock file with exact versions of all dependencies
• requirements.txt.backup - Backup of the original requirements.txt

Benefits of UV

• ⚡ Fast: Much faster than pip for dependency resolution and installation
• 🔒 Reliable: Lockfile ensures reproducible builds across environments
• 📦 Modern: Uses standard pyproject.toml configuration
• 🛠️ All-in-one: Handles virtual environments, dependencies, and project management

⚙️ AI Capabilities

• Uses LLM with RAG to:
  • Intelligently extract relevant resume sections for each question
  • Semantically match skills and experience to job requirements
  • Extract job-specific skills
  • Replace outdated resume skills
  • Tailor and regenerate resume (DOCX to PDF) (experimental)
  • Answer custom LinkedIn application questions with focused context
  • Evaluate job fit with relevant experience matching
  • Compress context for optimal small LLM performance

🎯 RAG Performance Optimizations

For Small LLMs (phi4-mini, TinyLlama, etc.)

• Context limited to 1500-2000 tokens for optimal performance
• Semantic relevance scoring ensures only pertinent information is included
• Section-aware chunking maintains context coherence
• Caching reduces repeated computations

Debug Mode

Enable debug mode to see RAG in action:

# Shows relevance scores and selected chunks
ai_generator = AIResponseGenerator(..., debug=True)

🔮 Future Work

• Expand support for more external job platforms
• Add a user-friendly GUI for configuration and monitoring
• Enhance error handling and reporting
• Community-driven plugin system for new features
• Advanced RAG features: Multi-modal embeddings, query expansion, hybrid search

📁 Repository Status

This project is a modified version of a popular LinkedIn Easy Apply Bot with enhanced AI capabilities via offline models and advanced RAG implementation.

Original credit: https://github.com/NathanDuma/LinkedIn-Easy-Apply-Bot

📜 License

This project is for educational and personal use only. Do not use it to spam applications or violate LinkedIn's terms.

---

🔐 Important: Keep your config.yaml and credentials private. Do not upload them to any public repo.