AccessibleAI - Multi-Agent System for Digital Content Accessibility

Kaggle 5-Day AI Agents Intensive - Capstone Project
Track: Agents for Good
Status: ✅ Production Ready

---

🎯 Quick Links

• GitHub Repository: https://github.com/Swanand33/accessible-ai-agents
• Main Implementation: /adk_version/agent.py
• Full Capabilities: /CAPABILITIES.md
• Kaggle Submission: /KAGGLE_FINAL_SUBMISSION.md

---

📖 Table of Contents

• Problem Statement
• Solution
• Capabilities Demonstrated
• Quick Start
• Architecture
• Installation
• Usage
• Testing
• Deployment
• Contributing
• License

---

🚨 Problem Statement

The Accessibility Crisis

• 2.2 billion people worldwide have vision impairment or blindness
• Only 3% of web images have meaningful alt-text descriptions
• 95% of PDFs are completely inaccessible to screen readers
• This excludes people with disabilities from:
  • 📚 Education: Textbooks, research papers, course materials
  • 💼 Employment: Job applications, work documents, training materials
  • 📰 Information: News, government documents, healthcare information

Impact

Without accessible digital content, people with visual impairments face:

• Barriers to education and career advancement
• Inability to access critical information independently
• Dependence on others for basic tasks
• Exclusion from digital society

---

💡 Solution

AccessibleAI is a production-ready multi-agent system built with Google's Agent Development Kit (ADK) that automatically makes digital content accessible by:

1. Generating descriptive alt-text for images using AI vision
2. Extracting and structuring text from PDF documents
3. Orchestrating specialized agents to handle different content types seamlessly

Why Multi-Agent Architecture?

• Specialization: Each agent optimized for its specific task
• Scalability: Easy to add new content types (audio, video, HTML, etc.)
• Reliability: Failure in one agent doesn't crash the entire system
• Observability: Track and log each agent's performance independently
• Maintainability: Clear separation of concerns and responsibilities

---

✨ Features

🔥 Core Features

1. Multi-Agent Architecture ⭐

   • Coordinator Agent (orchestration & routing)
   • Image Description Agent (Gemini Vision)
   • PDF Processing Agent (PyPDF2)

2. Tools Integration 🛠️

   • Google Gemini 2.0 Flash API (vision AI)
   • PyPDF2 (PDF text extraction)
   • Pillow (image processing)

3. Comprehensive Observability 📊

   • Structured logging for all operations
   • Success/failure tracking
   • Performance metrics & statistics

🎁 Bonus Features

• Batch Processing: Process multiple files at once
• Flexible Detail Levels: Concise or detailed descriptions
• Error Handling: Graceful handling of corrupt/invalid files
• Summary Generation: Human-readable processing reports
• Demo Mode: Works without API key for testing

---

📊 Capabilities Demonstrated

We demonstrate 6 core capabilities from the course (exceeds 3+ requirement):

#	Capability	Evidence	Day
1	Google ADK Framework	Full ADK implementation with 3 agents	Days 1-5
2	Multi-Agent Orchestration	Coordinator pattern for agent collaboration	Day 1
3	Tools & Integration	Gemini Vision API + PyPDF2	Day 2
4	Error Handling	Comprehensive error management	Day 4
5	Quality & Testing	6 tests, 100% pass rate	Day 4
6	Production Readiness	Tested, documented, deployable	Day 5

Total: 6 capabilities ✓ (Exceeds 3+ requirement)

See /CAPABILITIES.md for detailed breakdown.

---

🏗️ Architecture

System Overview

┌─────────────────────────────────────────────────┐
│           COORDINATOR AGENT                     │
│       (Orchestrates workflow)                   │
└─────────────┬───────────────────────────────────┘
              │
    ┌─────────┴──────────┐
    │                    │
    ▼                    ▼
┌──────────────┐  ┌──────────────┐
│ Image Agent  │  │ PDF Agent    │
│(Gemini Vision│  │(PyPDF2)      │
│ API)         │  │              │
└──────────────┘  └──────────────┘

Agent Responsibilities

1. CoordinatorAgent (Root Orchestrator)

• File type detection and routing
• Result aggregation
• Batch processing support
• Error handling and fault isolation

2. ImageDescriptionAgent (Vision AI)

• Analyzes images using Gemini Vision
• Generates descriptive alt-text
• Concise & detailed description modes

3. PDFProcessingAgent (Document Processing)

• Extracts text from PDFs
• Preserves document structure
• Handles multi-page documents

See /ARCHITECTURE_DIAGRAM.md for detailed diagrams.

---

🚀 Quick Start

Prerequisites

• Python 3.10 or higher
• Google Gemini API key (Get one here)

1. Clone Repository

git clone https://github.com/Swanand33/accessible-ai-agents.git
cd accessible-ai-agents

2. Create Virtual Environment

# Create
python -m venv agent-env

# Activate (Windows)
agent-env\Scripts\activate

# Activate (Mac/Linux)
source agent-env/bin/activate

3. Install Dependencies

pip install --upgrade pip
pip install -r requirements.txt

4. Configure API Key

# Copy example file
cp .env.example .env

# Edit .env and add your API key
# GEMINI_API_KEY=your_actual_api_key_here

5. Run Demo

# Run with ADK (requires ADK CLI)
adk run

# Or test locally
python adk_version/test_adk_agents.py

---

💻 Usage

Basic Python Usage

import google.generativeai as genai
from adk_version.agent import generate_image_description_tool, extract_pdf_text_tool

# Configure API
genai.configure(api_key="your_api_key")

# Process an image
result = generate_image_description_tool("path/to/image.jpg", detail_level="concise")
if result['success']:
    print(f"Alt-text: {result['alt_text']}")

# Process a PDF
result = extract_pdf_text_tool("path/to/document.pdf")
if result['success']:
    print(f"Extracted {result['char_count']} characters from {result['page_count']} pages")

Batch Processing

# Process multiple files
files = ["image1.jpg", "image2.png", "document.pdf"]

results = {
    "total": len(files),
    "processed": [],
    "failed": []
}

for file_path in files:
    if file_path.lower().endswith(('.jpg', '.jpeg', '.png', '.gif', '.webp')):
        result = generate_image_description_tool(file_path)
    elif file_path.lower().endswith('.pdf'):
        result = extract_pdf_text_tool(file_path)
    
    if result['success']:
        results["processed"].append(file_path)
    else:
        results["failed"].append(file_path)

print(f"Success: {len(results['processed'])}/{len(files)}")

---

🧪 Testing

Run All Tests

cd adk_version
python test_adk_agents.py

Test Coverage

• ✅ File type detection
• ✅ Image description generation
• ✅ PDF text extraction
• ✅ ADK agent initialization
• ✅ End-to-end image processing
• ✅ End-to-end PDF processing

Result: 6/6 tests passing (100%)

---

📁 Project Structure

accessible-ai-agents/
├── adk_version/                    # ADK Implementation (MAIN)
│   ├── agent.py                   # 3 ADK agents + tools
│   ├── agent.yaml                 # ADK deployment config
│   ├── test_adk_agents.py         # Test suite (6 tests)
│   ├── requirements.txt           # Dependencies
│   └── README.md                  # ADK-specific docs
├── src/                           # Original implementation
│   ├── agents/
│   │   ├── coordinator.py
│   │   ├── image_agent.py
│   │   └── pdf_agent.py
│   ├── config.py
│   └── utils/
├── tests/                         # Additional tests
├── examples/                      # Sample files
│   ├── sample_images/
│   └── sample_pdfs/
├── docs/                          # Documentation
├── CAPABILITIES.md                # Capabilities breakdown
├── ARCHITECTURE_DIAGRAM.md        # System diagrams
├── KAGGLE_FINAL_SUBMISSION.md    # Competition submission
├── README.md                      # Main documentation
├── requirements.txt               # All dependencies
├── .env.example                  # Environment template
├── .gitignore                    # Git ignore rules
└── LICENSE                       # MIT License

---

🚀 Deployment

Google Cloud (Agent Engine)

# Navigate to ADK version
cd adk_version

# Deploy with ADK
adk deploy --project-id=YOUR_PROJECT_ID --region=us-central1

# Check deployment
adk info

# Access web interface
adk web --port 8000

Local Development

# Run ADK locally
adk run

# Run tests locally
python test_adk_agents.py

# Use as library
python
>>> from agent import generate_image_description_tool
>>> result = generate_image_description_tool("image.jpg")

---

🎓 Real-World Impact

Use Cases

1. Educational Institutions 🎓

   • Make course materials accessible (ADA compliance)
   • Convert textbooks for all students
   • Support diverse learning needs

2. Digital Libraries 📚

   • Convert archives to accessible formats
   • Preserve historical documents accessibly
   • Enable researchers with disabilities

3. E-commerce 🛒

   • Product images accessible to all users
   • Improve customer experience
   • Expand market reach

4. Government 🏛️

   • Meet legal accessibility requirements
   • Serve constituents with disabilities
   • Ensure equal access to services

5. Businesses 💼

   • Comply with accessibility laws (ADA, WCAG)
   • Reduce legal liability
   • Expand customer base

Impact Numbers

• Manual alt-text: 2-5 minutes per image
• Our system: Seconds per image
• Cost: $50-100/hour manual vs. $0.001 per image
• 100x faster and more cost-effective
• Reach: 2.2 billion people with vision impairment

---

🛠️ Technical Excellence

Code Quality

• ✅ Type hints on all functions
• ✅ Comprehensive docstrings
• ✅ Error handling with fallbacks
• ✅ Input validation
• ✅ Structured return values

ADK Best Practices

• ✅ Proper agent configuration
• ✅ Tool registration and schemas
• ✅ Deployment configuration (agent.yaml)
• ✅ Environment management
• ✅ Production-ready structure

Testing & Documentation

• ✅ 6 comprehensive tests (100% pass rate)
• ✅ README with examples
• ✅ API documentation
• ✅ Capabilities mapping
• ✅ Architecture diagrams

---

📊 Competition Compliance

Requirement	Status	Evidence
Uses Google ADK	✅ YES	adk_version/agent.py + agent.yaml
3+ Capabilities	✅ YES	6 capabilities (see CAPABILITIES.md)
Multi-agent System	✅ YES	3 specialized agents
Solves Real Problem	✅ YES	Accessibility for 2.2B people
Working Implementation	✅ YES	6/6 tests passing
Documentation	✅ YES	Comprehensive README + guides
Production-Ready	✅ YES	Tests, deployment config, error handling

Overall Compliance: 7/7 (100%) ✅

---

🤝 Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create feature branch (git checkout -b feature/AmazingFeature)
3. Commit changes (git commit -m 'Add some AmazingFeature')
4. Push to branch (git push origin feature/AmazingFeature)
5. Open Pull Request

---

📄 License

This project is licensed under the MIT License - see LICENSE file for details.

---

🙏 Acknowledgments

• Google Gemini Team for the powerful vision API
• PyPDF2 Contributors for PDF processing capabilities
• Accessibility Community for highlighting critical needs
• Kaggle for the 5-Day AI Agents Intensive Course
• Course Instructors for excellent teaching

---

📞 Contact & Support

• GitHub Issues: Report bugs or request features
• GitHub Discussions: Ask questions or discuss ideas

---

🎯 What's Next?

Future Enhancements

The multi-agent architecture makes it easy to add:

• Video Captioning Agent - Automatic subtitle generation
• Audio Transcription Agent - Speech-to-text conversion
• HTML Optimization Agent - Web accessibility improvements
• Memory Capability - Remember user preferences
• Evaluation Framework - Quality metrics

Roadmap

• Add video support
• Add audio transcription
• Web API interface
• Dashboard for monitoring
• Scaling for enterprise use

---

Built with ❤️ for accessibility. Powered by Google Gemini 2.0 Flash.

Making digital content accessible for everyone. One file at a time.