🌍 Inspiration
Biosphere 2 stands as a testament to our desire to understand Earth’s intricate ecosystems and the potential of off-planet living. We were inspired by the challenge of combining environmental science and artificial intelligence to simulate a self-sustaining ecosystem navigator. Our mission: to empower scientific discovery using AI-powered digital twins.
🤖 What it does
B2Twin is an AI Digital Twin Navigator that:
- Ingests and analyzes real-time Biosphere 2 sensor data.
- Uses local LLMs to generate scientific insights for ecosystem health.
- Trains ML models on individual environmental zones.
- Tracks crop health, monitors anomalies, and suggests corrective actions.
- Enables inter-agent communication across AI systems via port-based protocols.
- Simulates collaborative decision-making using a secondary assistant AI (Small Talk Agent).
🔧 How we built it
- Streamlit UI for interactive visualization and simulation.
- Ollama + Gemma3:4B LLM for local large language model analysis.
- Scikit-learn ML models for zone-level predictive learning.
- REST API (Flask) for inter-AI communication across local ports.
- Prompt engineering to simulate multi-agent scientific collaboration.
- Custom CSV uploader + DataCleaner to ingest and clean diverse datasets dynamically.
🚧 Challenges we ran into
- Handling inconsistent metadata and unit conversions across 24+ datasets.
- Balancing ML/LLM inference speed with system responsiveness.
- Creating a lightweight architecture that runs fully offline on local devices.
- Integrating multi-agent communication and memory tracking while maintaining modularity.
🏆 Accomplishments that we're proud of
- Built a full-featured AI-powered digital twin from scratch within 48 hours.
- Seamless integration of ML + LLM + REST API in a scientific application.
- Developed a multi-agent ecosystem capable of simulating inter-AI conversations.
- Created a scalable foundation for digital twin research and future collaborations.
📚 What we learned
- How to integrate AI agents across modalities (ML + LLM) and communication layers.
- Real-world prompt engineering for scientific reasoning tasks.
- Modular system design with reusable data pipelines and visualization components.
- The power of local LLMs in domain-specific ecological simulations.
🚀 What's next for Hack Arizona - devCon4
- Enhance multi-agent memory and knowledge-sharing systems.
- Connect with real-time cloud Biosphere 2 data feeds.
- Incorporate XR interfaces for immersive ecosystem monitoring.
- Publish this as an open-source AI framework for scientific digital twin simulation.
- Extend this architecture for off-planet terraforming simulation and sustainability research.
Log in or sign up for Devpost to join the conversation.