Making Machines Plan - Disaster Response

• 

  MMP.AI Logo

• 

  Example Recommended Actions Plan

• 

  Report Start, with situation summary and a number of areas of interest marked on maps

• 

  The statistics for a given run, showing that the AI has used 52 tools and thus re-prompted itself 50 times to explore an optimal solution

Inspiration

The increasing frequency of natural disasters and emergencies highlighted the need for a smart tool that could support first responders in making quick, informed decisions, enhancing both preparedness and real-time response efforts.

What it does

MMP.AI assists first responders by generating both pre-emptive and reactive disaster response plans. Using AI-driven chain-of-thought reasoning, it integrates real-time data to create adaptive, situation-specific strategies.

How we built it

I built MMP.AI with a ground-up approach (I would have self-hosted the models as well if I had the GPU). I then built a custom "o1 style" reasoning model on-top of an existing LLM before self-defining and providing a tools architecture such that the model can decompose a probem before calling tools such as map generation and local area expertise to effectively solve the problem.

Challenges we ran into

LLMs have a tendency to be tempremental, especially when you are prompting them to do very non-LLM things (such as calling external programs and interpreting its own logical plans).

Accomplishments that we're proud of

We’re proud of creating an intelligent system capable of thinking through complex disaster scenarios and delivering actionable, customized plans in real-time. The integration with external systems was a major achievement. The system does actually work! (with some hallucinations because I didn't have time to add a geocoding tool) and I genuinely think could be useful if given some more time in the oven as I've seen the performance of this tool increase exponentially with the number of tools it is given access to

What we learned

I'm now thoroughly experienced with the concept of how LLMs use Tools, I think I've built quite a strong framework for handling tools in a simple way (if we remove LangChain from the equation) so I now feel more confident about building/learning about AI Agents in the future

What's next for MMP.AI

More Tools! Specificially more specialised tools for specific weather events, ideally having the majority of "Expert" tools having a large background of RAG knowledge to assist in a more accurate and representive way/

Built With

• python