Castanea

Three AI agents working in sync to research, analyze, and create — freeing you to focus on ideas that matter.

Comment

About the project

Castanea was inspired by the real challenges students face when working on academic tasks — from researching complex topics to summarizing dense texts and crafting well-structured essays. Too often, these tasks require switching between multiple tools, sources, and formats. I wanted to create a single AI-powered platform that could handle all three seamlessly, freeing up time for learning rather than tool management.

The core idea was to simulate a small “AI office” with specialized agents:

  • 🧠 ResearcherAgent for in-depth topic exploration using a Retrieval-Augmented Generation (RAG) pipeline powered by Perplexity and Gemini Pro.
  • ⚡ AnalystAgent for lightning-fast text summarization and key point extraction using Gemini Flash.
  • ✍️ WriterAgent for generating original content and saving it directly to files.

How I built it

The frontend is built with Vite, React, and JavaScript for a responsive and modern interface.
The backend uses Python, Agent Development Kit (ADK), and integrates with Gemini Pro/Flash models alongside Perplexity’s Sonar API.

Architecture overview:

  1. User inputs a request.
  2. The central dispatcher determines the correct specialist agent.
  3. The selected agent executes its task:
    • ResearcherAgent → multi-step query decomposition → Perplexity search → Gemini Pro synthesis.
    • AnalystAgent → text analysis → concise summarization.
    • WriterAgent → original content generation → optional file save.

I also included .env configuration for secure API key management and separated frontend/backend environments for scalability.

What I learned

  • How to design an agent-based architecture that feels natural to the user.
  • The importance of clear task routing to avoid overloading a single AI model.
  • Practical integration of RAG workflows combining real-time search with LLM synthesis.
  • Fine-tuning prompt engineering for consistency across multiple AI agents.

Challenges I faced

  • Time constraints: Designing, coding, and integrating multiple agents within hackathon deadlines.
  • Balancing speed and accuracy: Finding the right trade-off between Gemini Pro’s depth and Gemini Flash’s speed.
  • Query decomposition: Teaching the ResearcherAgent to break broad topics into precise sub-questions for better results.
  • User experience: Making the interface intuitive while showcasing the power of three different agents.

Built With

Share this project:

Updates