CrisisForge

Multi-agent AI that outthinks human crisis teams.

Comment 2

Inspiration

Living in India, I've witnessed how devastating floods and cyclones can strike coastal areas with little warning. Rescue teams often struggle with overwhelming data, emotional pressure, and difficult trade-offs under extreme time constraints.

When the MLH AI Hackfest announced the theme "Build hacks that can think and take decisions even better than humans!", it felt like the perfect opportunity. I wanted to create an AI system that doesn't panic, doesn't get biased by limited information, and can evaluate hundreds of scenarios in seconds — something no human team could achieve in a real crisis.

What it does

CrisisForge is a multi-agent AI system that thinks and makes optimal decisions in high-stakes crises better than humans.

You describe a crisis (for example: “Severe flood in coastal Mumbai affecting 10,000 people, limited rescue boats, and an incoming cyclone”).

Four specialized agents then collaborate in real-time:

  • Data Scout gathers live weather, news, and location data
  • Risk Forecaster runs rapid simulations to predict outcomes
  • Action Optimizer evaluates dozens of possible plans and scores them
  • Ethical Guardian ensures decisions remain fair and unbiased

The system delivers a clear, step-by-step action plan with confidence scores, alternative options, and a natural voice briefing — all in seconds.

How we built it

How we built it

We built CrisisForge using a completely free tech stack to keep it accessible and hackathon-friendly.

  • Multi-agent orchestration: LangGraph (LangChain) to create four collaborating agents that can reason together, hand off tasks, and show transparent step-by-step thinking.
  • Core reasoning: Google Gemini 2.0 Flash through Google AI Studio — powerful free LLM with strong chain-of-thought capabilities for complex crisis decisions.
  • Frontend: Streamlit for a clean, interactive interface where users input a crisis and watch the agents debate and decide in real-time.
  • Real-time data: Open-Meteo (completely free weather API, no key needed) for flood/cyclone data + NewsAPI free tier for live updates.
  • Mapping: Leaflet.js for simple location visualization.
  • Deployment: Hosted for free on Streamlit Community Cloud — just linked our GitHub repo.

The agents are specialized:

  • Data Scout pulls live information
  • Risk Forecaster runs quick simulations (using simple Python Monte Carlo logic)
  • Action Optimizer scores multiple plans
  • Ethical Guardian adds fairness checks

Everything runs smoothly within free limits, and we added clear reasoning traces so anyone can see how the AI thinks and decides better than humans.

Challenges we ran into

The biggest challenge was making multiple AI agents collaborate effectively without hallucinating or going in loops. Tuning the prompts and adding clear hand-off rules took significant time and iteration.

We also faced rate limits with real-time APIs during testing, so we implemented simple caching as a fallback.

Integrating voice output smoothly while keeping the entire demo responsive under hackathon time pressure was another hurdle.

Despite these issues, breaking the system into specialized agents helped us manage complexity better than a single large prompt.

Accomplishments that we're proud of

  • Successfully built a working multi-agent system that demonstrates superhuman decision-making in under 24 hours
  • Created transparent reasoning traces so anyone can see how the AI thinks step-by-step
  • Integrated real-time data sources and voice output for a polished, impactful live demo
  • Built something directly aligned with the hackathon theme: AI that truly thinks and decides better than humans in critical situationsWe learned how powerful multi-agent architectures are compared to single LLM calls — breaking complex reasoning into specialized roles leads to much more reliable and thoughtful decisions.

We also discovered the importance of clear agent communication protocols and ethical guardrails in decision-making systems.

On the technical side, we gained hands-on experience with CrewAI, prompt engineering for collaboration, and integrating multiple APIs under time pressure.

Most importantly, we saw firsthand how AI can process far more data and scenarios than any human team, opening our eyes to the real potential of agentic AI in high-stakes environments.

  • Made the project fully functional with an intuitive interface that judges and users can interact with instantly

What we learned

We learned how powerful multi-agent architectures are compared to single LLM calls — breaking complex reasoning into specialized roles leads to much more reliable and thoughtful decisions.

We also discovered the importance of clear agent communication protocols and ethical guardrails in decision-making systems.

On the technical side, we gained hands-on experience with CrewAI, prompt engineering for collaboration, and integrating multiple APIs under time pressure.

Most importantly, we saw firsthand how AI can process far more data and scenarios than any human team, opening our eyes to the real potential of agentic AI in high-stakes environments.

What's next for CrisisForge

In the future, we plan to:

  • Add more domain-specific agents (healthcare, finance, logistics)
  • Integrate advanced simulation tools for even more accurate forecasting
  • Make the system mobile-friendly for on-ground use by rescue teams
  • Add support for voice input so users can describe crises naturally
  • Explore on-chain logging (using Solana) for transparent and auditable decisions in government or NGO use cases

We ultimately want CrisisForge to become a practical tool that helps real disaster management teams make faster and smarter decisions when lives are on the line.

Built With

Share this project:

Updates