Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

IgnisAI

Real-Time AI Navigation for Structural Firefighting

IgnisAI is a multi-agent AI system designed for GPS-denied indoor environments. It parses building floor plans, models live fire and smoke spread, detects survivors via thermal imaging, and delivers spoken turn-by-turn navigation to each firefighter's earpiece — updating every 30 seconds as conditions change. No screen required.


The Problem

Firefighters entering burning structures operate with incomplete information. Floor plans are rarely memorized, conditions change within seconds, and radio communication is unreliable under duress. Existing navigation tools depend on GPS, which is unavailable indoors. The result: delayed rescues, wrong turns, and preventable casualties.


The Solution

IgnisAI deploys five specialized AI agents that work in coordination to give each firefighter a real-time, spoken route — recalculated continuously as fire spreads and new information arrives.


Agent Architecture

Agent 1 — Floor Plan Parser Accepts a building blueprint image and uses Claude Vision to extract rooms, corridors, stairwells, and exits. Outputs a weighted NetworkX graph representing the building topology.

Agent 2 — Fire Spread Simulator Models fire and smoke propagation using a cellular automaton across the building graph. Ingests live sensor data and updates danger scores on each node every 30 seconds.

Agent 3 — Survivor Detector Runs YOLOv8 on a thermal camera feed to identify heat signatures consistent with human presence. Outputs survivor location as a node on the building graph.

Agent 4 — Route Planner Executes danger-weighted A* pathfinding on the current graph state. Nodes with danger scores above 0.8 are marked impassable. Routes are recalculated automatically when fire spread or survivor detection events occur.

Agent 5 — Voice Output Converts the active route into natural-language turn-by-turn instructions and delivers them via text-to-speech to the firefighter's earpiece. No screen interaction required.


Tech Stack

Layer Technology
Core backend Python 3
AI reasoning Claude API (Anthropic)
Fire simulation NumPy cellular automaton
Building graph NetworkX
Pathfinding A* (danger-weighted)
Survivor detection YOLOv8 (Ultralytics)
Image processing OpenCV
Voice output Web Speech API
Dashboard HTML5 Canvas, vanilla JS
Real-time sync WebSocket

Demo Scenario

  1. A building floor plan is parsed into a live graph with 12 nodes and 13 edges.
  2. Fire originates in Office 2. The danger map begins spreading across connected nodes.
  3. Two firefighter units are deployed. Agent 4 routes each unit around active danger zones.
  4. Fire spreads to Corridor A. Both routes are recalculated instantly.
  5. Agent 3 detects a thermal signature in Office 3. FF-1 Alpha is rerouted to perform the rescue.
  6. Agent 5 speaks each navigation step aloud. The entire sequence runs in under 90 seconds.

Project Structure

ignisai/
├── agents/
│   ├── agent1_floorplan.py     # Floor plan parser
│   ├── agent2_fire.py          # Fire spread simulator
│   ├── agent4_router.py        # Danger-weighted A* router
│   └── agent5_voice.py         # Voice instruction generator
├── data/
│   ├── graph.json              # Building graph
│   ├── danger.json             # Current danger state
│   ├── snapshots.json          # Fire spread history
│   └── routes.json             # Active firefighter routes
├── frontend/
│   └── index.html              # Live ops dashboard
└── main.py                     # Orchestrator

Running the Project

Install dependencies

python3 -m venv venv
source venv/bin/activate
pip3 install anthropic networkx numpy opencv-python pillow requests

Run agents

python3 agents/agent2_fire.py     # Generate fire spread data
python3 agents/agent4_router.py   # Calculate safe routes

Open dashboard

open frontend/index.html

The dashboard runs entirely in the browser. No server required.


Design Principles

GPS-denied by design. The system uses building topology, not satellite positioning. It works in basements, concrete structures, and signal-blocked environments.

Screen-free operation. All navigation is delivered as spoken audio. Firefighters keep their hands and attention free.

Continuous recalculation. Routes are not static. Every agent update triggers a new pathfinding pass. A blocked corridor is detected and rerouted within one cycle.

Modular agent design. Each agent is independently replaceable. The thermal detection agent can be swapped between simulated and live YOLOv8 inference without changing any other component.


Team

Built for the KXSB Hackathon — April 2026.

About

Multi-agent AI system that parses floor plans, simulates fire spread, detects survivors via thermal imaging, and delivers real-time spoken navigation to firefighters — no GPS, no screen required.

Topics

Resources

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages