• 

Inspiration

Natural disasters like floods, cyclones, and earthquakes often create chaos within minutes. During such situations, people panic, emergency lines become overloaded, and rescue teams struggle to prioritize cases effectively. The inspiration for CrisisMind AI came from observing how critical the first few minutes are during emergencies. I wanted to build a system where AI could respond instantly — reducing panic, organizing information, and supporting both victims and rescue teams in real time.

What it does

🎤Listens to voice or text messages from people during emergencies

🧠 Understands urgency and context using AI reasoning

🚨 Detects risk level (Low / Medium / High)

📍 Extracts important details like location, number of people, medical needs

🛟 Provides instant safety guidance to the victim

📊 Sends structured alerts to a rescue dashboard

⚡ Automatically prioritizes cases based on severity It turns panic messages into organized, actionable information so rescue teams can respond faster and more efficiently.It’s like a 24/7 intelligent emergency operator that never gets overloaded.

How we built it

1️⃣ Designing the Architecture

We first designed a cloud-native architecture focused on scalability and real-time processing. The system was structured into three main layers:

User Interaction Layer (Web Interface)

AI Processing Layer

Rescue Dashboard & Analytics Layer

2️⃣ Building the Frontend

We built the web interface using:

React.js

Tailwind CSS

The frontend includes:

Live chat interface

Voice input button

Panic level indicator

Real-time rescue dashboard

The UI was designed to be minimal and calm, suitable for emergency situations.

3️⃣ Implementing AI Processing

For intelligence and reasoning, we used:

Gemini via

Vertex AI

Gemini processes incoming voice or text messages and:

Detects urgency

Identifies vulnerable individuals (elderly, injured)

Extracts structured information

Generates safety recommendations

We used prompt engineering to ensure the output was structured in JSON format for dashboard integration.

4️⃣ Voice Integration

To support real-time voice interaction, we integrated:

Google Cloud Speech-to-Text

Google Cloud Text-to-Speech

This allows the system to:

Convert voice to text

Process it using Gemini

Respond back in natural voice

5️⃣ Real-Time Dashboard

We used:

Firebase for live data updates

Structured alert cards showing:

Location

Risk level

Number of people

Medical priority

Each emergency case is ranked automatically using a weighted risk scoring formula.

6️⃣ Risk Scoring Logic

We implemented a scoring system:

𝑅 𝑖 𝑠 𝑘 𝑆 𝑐 𝑜 𝑟

𝑒

( 0.5 × 𝑈 𝑟 𝑔 𝑒 𝑛 𝑐 𝑦 ) + ( 0.3 × 𝑉 𝑢 𝑙 𝑛 𝑒 𝑟 𝑎 𝑏 𝑖 𝑙 𝑖 𝑡 𝑦 ) + ( 0.2 × 𝐿 𝑜 𝑐 𝑎 𝑡 𝑖 𝑜 𝑛 𝑆 𝑒 𝑣 𝑒 𝑟 𝑖 𝑡 𝑦 ) RiskScore=(0.5×Urgency)+(0.3×Vulnerability)+(0.2×LocationSeverity)

This helps prioritize rescue cases intelligently.

7️⃣ Deployment

Frontend deployed using Firebase Hosting

Backend deployed on Google Cloud

Environment variables secured for API access

Designed to scale automatically using cloud infrastructure

🚀 Final Outcome

The result is a fully functional AI-powered live disaster agent that:

Responds in real time

Converts unstructured panic into structured intelligence

Prioritizes rescue operations automatically

Scales for large disaster scenarios

Challenges we ran into

1️⃣ Handling Unstructured Panic Input

During disasters, people don’t speak in clear sentences. Messages are often incomplete, emotional, or chaotic.

Designing prompts that could reliably extract structured data (location, urgency, medical needs) from messy input was a major challenge.

We solved this using structured prompt engineering and enforced JSON output formatting.

2️⃣ Real-Time Latency

Emergency systems must respond instantly.

Balancing:

Voice processing

AI reasoning

Structured output generation

while keeping response time low required optimizing API calls and reducing unnecessary processing steps.

3️⃣ Designing a Fair Risk Scoring Model

Creating an automated priority system was difficult.

We had to carefully balance factors like:

Urgency level

Vulnerable individuals (elderly, injured)

Location severity

If weights were too high or too low, prioritization would become biased or inaccurate.

4️⃣ Ensuring Scalable Architecture

Disasters can generate thousands of simultaneous requests.

Designing a cloud-native system that could scale without crashing required thoughtful architecture planning using managed cloud services.

5️⃣ Building a Calm User Experience

In high-stress situations, a complex interface increases panic.

We had to:

Simplify UI elements

Reduce cognitive load

Avoid overwhelming animations

Make information easy to understand instantly

Designing for stress was more challenging than designing for normal users.

6️⃣ Structured AI Output Reliability

Large language models sometimes generate inconsistent output formats.

Ensuring predictable, machine-readable responses required:

Strict output templates

Validation checks

Fallback logic

🚀 What These Challenges Taught Us

Real-world AI is not just about accuracy — it’s about reliability.

Human-centered design is critical in emergency systems.

Scalability must be planned from day one.

Structured intelligence is more powerful than raw AI responses.

Accomplishments that we're proud of

Built a real-time AI disaster response live agent that converts voice/text into structured emergency alerts.

Designed an intelligent risk scoring model for automatic rescue prioritization.

Integrated Gemini via Vertex AI** with a scalable cloud-native architecture.

Developed a live rescue dashboard using Firebase.

Created a human-centered, stress-friendly interface for emergency situations.

What we learned

Building real-world AI requires reliability, not just intelligence.

Structuring LLM outputs into actionable data is critical for practical impact.

Cloud-native design is essential for scalability during high-demand situations.

Human-centered UI design matters even more in high-stress environments.

Collaboration between AI reasoning and system architecture creates meaningful solutions.

What's next for CrisisSphere AI

🔄 Integrate real-time weather and disaster APIs for predictive alerts.

📍 Add automatic GPS location detection for faster rescue coordination.

🌍 Expand multilingual support for rural and global deployment.

🤝 Partner with local authorities and NGOs for pilot testing.

📊 Enhance the risk model using historical disaster data and machine learning.

☁️ Scale infrastructure for large-scale, real-world emergency scenarios.

Built With

• javascript
• python