FirstWaveAI

System Architecture
Agentic AI System

💡 Inspiration

📞 Over 240 million emergency calls are made in the United States each year (National Emergency Number Association), putting extreme pressure on dispatchers in life-or-death situations. In those critical first moments, key details can be missed, delayed, or misunderstood when callers are panicking.

FirstWaveAI was built to rethink the emergency intake process. I created an AI-powered call assistant that speaks directly with callers, asks clarifying questions, and structures critical information in real time. The system then visualizes the situation on a live map, identifies nearby resources, and generates an AI-assisted dispatch recommendation, all while keeping a human dispatcher in full control with an approval override. 🚑

🚨 What it does

FirstWaveAI is a real-time emergency dispatch assistant that combines speech recognition, multi-agent AI, and interactive visualization to help dispatchers work faster and more accurately.

🔧 Core Features

🎙️ Voice-First Interface

Callers can speak naturally using the Web Speech API, while the system transcribes the conversation in real time and maintains a full transcript.

🧠 Multi-Agent AI Pipeline (LangGraph + LLaMA 3.3 70B)

Six specialized AI agents work together to analyze the call:

📝 Extraction Agent – Captures key details (location, injuries, hazards, people count)
🚦 Triage Agent – Assigns priority levels (P1–P4)
❓ Next-Question Agent – Suggests clarifying follow-ups
🚓 Dispatch Planner – Recommends EMS, Fire, or Police
🗺️ Resource Locator – Finds nearest available units with ETAs
🛡️ Safety Guardrail – Ensures ethical RECOMMENDATIONS

🖥️ Interactive Dashboard

A clean three-column interface shows:

Live chat transcript 💬
AI-generated emergency summary 📝
Dispatch recommendations with approve/cancel controls ✅❌

🗺️ Resource Mapping

An interactive Leaflet map displays nearby hospitals, fire stations, police, and pharmacies with distances and travel times.

🛠️ How I built it

Frontend

Next.js 16 + React 19
Tailwind CSS 4 (custom emergency theme)
shadcn/ui components
Leaflet + Openstreetmap
Web Speech API
TypeScript

Backend

FastAPI
LangGraph (multi-agent orchestration)
Groq + LLaMA 3.3 70B
Fish Audio API (TTS)
Server-Sent Events (SSE) for real-time updates

⚠️ Challenges I ran into

Building AI for real emergencies is hard.

🔁 Avoiding redundant questions – The AI kept asking things the caller already said, so I had to add strict memory rules.
⏱️ Speed vs. accuracy – P1/P2 emergencies required immediate action, while P3/P4 allowed more questioning. Encoding this logic took serious prompt tuning.
⚖️ Ethical safety – I carefully designed guardrails to prevent harmful questions.

🏆 Accomplishments I’m proud of

🎯 This was my first fully solo hackathon project!

📚 What I learned

🗂️ Solo project management — better scoping, prioritization, and focus
🐞 Debugging under pressure — isolating issues across multiple systems quickly