Wildfire Detection and Relief w/ Agents+Cloud ML inference

Inspiration

Every year, wildfires destroy millions of acres, displace thousands of people, and cause untold ecological and economic damage. Current detection systems often react after the damage has begun. I wanted to change that — to create an intelligent, automated, end-to-end disaster relief agent network that not only detects wildfires in real time but also coordinates immediate response.

Inspired by how multi-agent AI and LLM orchestration systems (like AWS AgentCore and Google’s Gemini) are transforming industries, I asked myself:

“What if agents could monitor live data, analyze anomalies, and instantly generate action plans — all autonomously?”

What I Built

My project is an wildfire relief orchestration system that uses agents to perform each stage of disaster management:

1. Detection Agent – Continuously monitors wildfire data streams (weather, thermal, satellite) and flags anomalies using models AWS Sagemaker.
2. Context Agent – Gathers nearby fire stations, evacuation zones, wind patterns, and previous similar cases from AWS DynamoDB.
3. Analysis Agent – Compare the situation with similar historical cases from DynamoDB and predicts risk escalation by analyzing previous response patterns, climate conditions and resource availability.
4. Action Agent – Generates a context-aware plan of action and triggers mock alerts via to local authorities and relief teams.
5. Visualization Dashboard – Built on Streamlit, showing real-time data, alerts, and recommended actions.

⚙️ How I Built It

I structured the system as modular agents connected through AgentCore and managed via MCP strands (for composable workflows):

• Data Ingestion: AWS Lambda pulls real-time data from wildfire APIs and weather sources into S3.
• Processing & Storage: Data is structured into JSON-like objects for DynamoDB, enabling efficient lookups.
• Agent Reasoning: Gemini and Bedrock models perform reasoning, similarity matching, and action generation.
• Communication: AWS SNS automates outbound alerts; Step Functions orchestrate multi-agent workflows.
• Visualization: Streamlit dashboard displays the full pipeline, showing which agents are active and their outputs.

💡 What I Learned

• How multi-agent frameworks like MCP and AgentCore can modularize complex systems.
• Integration of AWS Bedrock and Gemini to blend reasoning, retrieval, and execution.
• Importance of designing data pipelines for real-time monitoring and fail-safe response.
• How to build scalable prototypes quickly using AWS Step Functions + DynamoDB.

Challenges I Faced

• Managing data synchronization between multiple microservices in near real-time and most importantly finding a way to integrate gemini models, and using pre-trained models through AWS!
• Ensuring agents communicated efficiently without overflowing/too much context, summarizing the context that is passed by efficiently.
• Balancing hackathon time constraints with complex AWS configurations, and building with multiple AI agents and balancing code speed with retaining knowledge of how the code and internal systems work.
• Structuring data efficiently without relying on a vector DB due to time constraints, using clever DynamoDB schema design instead with a heuristic "vector" like column which summarized the data well enough .

Impact

This project demonstrates how cloud AI Inference through services like Sagemaker can be leveraged to solve high-stakes real-world problems.

Built With

• agentcore
• mlstate-opa
• sagemaker
• strands