1. What it does

Right now, LiveShield AI demonstrates how an AI agent can:

-Process disaster-related queries.

-Pull live data from sources like weather APIs and disaster reports.

-Generate helpful responses tailored to a user’s situation.

-It’s a working proof-of-concept that shows how an AI agent can transform scattered data into clear disaster guidance.

2. How we built it

-Fine-tuned the IBM Granite-13B Instruct LLM using QLoRA adapters to handle disaster-specific instructions.

-Implemented a simplified Plan → Act → Observe loop so the agent can break queries into steps and call the right tools.

-Integrated APIs like OpenWeatherMap and ReliefWeb for live updates.

-Built a Streamlit frontend to demonstrate the workflow end-to-end.

3. Challenges we ran into

-Running such a large LLM on limited Colab resources.

-Ensuring valid and structured outputs from the agent.

-Handling API rate limits and errors during testing.

-Managing the scope so we could build a usable demo within hackathon time.

4. Accomplishments that we're proud of

-Built an initial working version of the disaster management agent in a short time.

-Successfully combined a fine-tuned LLM with live APIs for context-aware responses.

-Delivered a simple but effective web-based interface for demo purposes.

-Created a project that is both technically promising and socially impactful.

5. What we learned

-How to use efficient fine-tuning techniques (LoRA/QLoRA) for big models.

-The importance of structuring LLM-agent reasoning to avoid unstructured answers.

-How real-time APIs make AI outputs more trustworthy and actionable.

-That even an MVP can demonstrate the future potential of AI in disaster management.

6. What’s next for LiveShield AI

-Our hackathon build is just the first step. If selected, we want to:

-Add more real-time data sources (traffic, satellite imagery, social media monitoring).

-Improve multi-language support for accessibility in different regions.

-Work with emergency agencies to tailor responses for real-world use.

-Scale the prototype into a full disaster response platform that can save liv

Built With

• bitsandbytes
• colab
• github
• granite
• hugging-face
• lora
• openstreetmap
• openweathermap
• peft
• python
• reliefweb
• streamlit