LearnForge AI

🎯 Project Story

Inspiration

Education is becoming increasingly complex, with learners needing to synthesize information from multiple sources - academic papers, videos, code repositories, and web articles. We built LearnForge AI to automate this research process using Google's Agent Development Kit (ADK) and multi-agent orchestration patterns.

What it does

LearnForge AI is an intelligent research platform that:

• Gathers sources from 5+ platforms (Google Search, Exa academic papers, Tavily web search, GitHub code, YouTube videos)
• Orchestrates 5 specialized AI agents using all 6 ADK patterns (Hierarchical, Parallel, Sequential, Loop, Generator-Critic, LLM)
• Extracts evidence and validates information using parallel processing
• Generates educational content including summaries and interactive quizzes
• Provides a REST API and web UI for seamless integration

How we built it

Architecture:

• Backend: Python with Google ADK 1.16.0
• Agent Patterns: Hierarchical coordinator, parallel source gathering, sequential processing, loop-based validation
• Tools Integration: MCP (Model Context Protocol) for third-party tools (Exa, Tavily, Firecrawl, GitHub)
• API Layer: Flask REST API with CORS support
• Frontend: Next.js 15 with TypeScript, Tailwind CSS, shadcn/ui
• Deployment: Google Cloud Run (serverless, auto-scaling)

Agent System:

1. Coordinator Agent (Hierarchical) - Routes queries to specialized sub-agents
2. Source Gatherer Agent (Parallel) - Simultaneously searches 5+ sources
3. Evidence Extractor Agent (Parallel) - Extracts key information from sources
4. Writer-Critic Agent (Generator-Critic) - Generates and validates research reports
5. Quiz Generator Agent (Loop) - Creates and validates educational quizzes

Tools Used:

• Google Gemini 2.0 Flash (LLM model)
• Google Search API
• Exa AI (academic papers)
• Tavily (web search)
• GitHub API (code examples)
• YouTube Data API v3 + Transcript API
• Firecrawl (content extraction)

Challenges we ran faced

1. Module Import Issues on Cloud Run - Resolved by creating proper package structure with agents/config.py and agents/requirements.txt
2. CORS Blocking Frontend - Fixed with explicit CORS configuration for localhost:3000
3. ADK Result Extraction - Properly extracted from result.events instead of direct iteration
4. MCP Server Integration - Successfully integrated 5+ third-party tools using Model Context Protocol
5. Deployment Configuration - Used adk deploy cloud_run --with_ui for proper UI deployment

Accomplishments that we're proud of

✅ Implemented all 6 ADK agent patterns in a production system ✅ Integrated 5+ real tool APIs (Exa, Tavily, GitHub, YouTube, Google Search) ✅ Deployed to Google Cloud Run with auto-scaling and serverless architecture ✅ Built complete REST API with health checks and proper error handling ✅ Created responsive Next.js frontend with real-time agent interaction ✅ Achieved sub-30 second research across multiple sources

What we learned

• Advanced multi-agent orchestration using Google ADK
• MCP (Model Context Protocol) for flexible tool integration
• Proper Cloud Run deployment with ADK CLI
• Managing complex async agent workflows
• Error handling and graceful degradation for missing dependencies
• Frontend-backend integration with CORS and proper API design

What's next for LearnForge AI

• Enhanced Tool Integration: Add more MCP servers (Notion, Slack, databases)
• Advanced Memory: Implement persistent session storage with Vertex AI Agent Engine
• Real-time Collaboration: Add WebSocket support for live research sessions
• Mobile App: Build native mobile applications
• Enterprise Features: Team workspaces, shared research libraries, analytics
• AI Model Selection: Allow users to choose between Gemini 2.0 Flash, 2.5 Pro, etc.

Built With

• adk
• exa
• exa-ai
• google-adk
• mcp
• python
• tavily