Prometheus

Beacon in Firefighter suit
Architecture Diagram
PCB Design (Inside Beacon)
Dashboard View
Dashboard Building Info
AI Command Advisor

Inspiration

Firefighters are the backbone of community safety, performing essential work that is as dangerous as it is under-appreciated. Every day, they step into environments where split-second decisions determine survival, yet they often do so relying on equipment that hasn't kept pace with modern technological leaps. The Bison Hackathon is a call to leverage AI for Truth and Service. Our team fulfills this goal by empowering the heroes who serve us through the integration of high-performance edge hardware and real-time data analysis.

What it does

Prometheus is a vision to develop the next generation of fire-fighter suits that will make our firefighters more efficient and help save more lives. We have come up with our own in-house hardware called Beacon, along with a thermal intelligence dashboard for firefighter operations. It displays normal (RGB) and thermal camera feeds side-by-side with live AI analysis.

The system provides:

Person Detection: Uses YOLOv8 and ByteTrack on the RGB feed to identify victims with stable tracking IDs.
Heat Zone Mapping: Identifies the top 5% hottest pixels to map hazards by direction (Left/Ahead/Right) and distance.
Flashover Risk Engine: Computes a real-time risk score (0–100); a HIGH risk score triggers an on-screen evacuation warning.
Command Advisor: Sends live telemetry (victim count, risk, heat zones) to a local LLM (Gemma 3:4B via Ollama) to generate tactical recommendations.

How we built it

We approached this as a two-front engineering challenge, combining rugged physical design with an advanced software pipeline:

The Hardware (Beacon): A compact 105mm × 85mm × 45mm tactical enclosure designed for center-chest mounting to ensure maximum stability, consistent camera alignment, and reduced motion artifacts during operation.
Thermal Protection: To protect the beacon from extreme heat exposure, we implemented a layered “thermal sandwich” composed of Aerogel insulation and Phase Change Materials (PCM). This passive thermal management system maintains safe internal operating temperatures while external environments range from 500°C to 1000°C.
Machine Learning (ML) Models: We deployed YOLOv8, using variants ranging from approximately 3–68 million parameters depending on performance requirements, for real-time object detection and classification. MiDaS is used for monocular depth estimation (not object detection) to infer spatial distance from a single RGB camera stream. All ML inference runs entirely locally on hardware powered by NVIDIA’s Jetson Nano, enabling low-latency, offline operation.
The Software (Prometheus): A Streamlit-based dashboard that processes dual-feed video locally in real time. Depth maps generated via MiDaS are converted from pixel space into actionable distance measurements in feet, supporting an effective operational range of 10–250 ft.
Edge Reliability: All perception, inference, and visualization pipelines execute locally using Ollama and native model weights, ensuring 100% system availability in communication-denied or infrastructure-compromised “blackout” environments where cloud-based AI solutions would fail.

Challenges we ran into

Our toughest challenge was what we came to call the Thermos Paradox. While insulating the device from external heat was effective, it also trapped the heat generated by the onboard processor.

To overcome this, we designed a custom thermal control system that adapts to the firefighter’s surroundings. When conditions permit, an internal fan and heatsink actively vent stored heat, effectively resetting the device. This approach allowed us to balance thermal protection and performance without sacrificing reliability during critical operations.

Accomplishments that we're proud of

One of our biggest accomplishments was demonstrating that powerful, real-time AI can run entirely at the edge. Prometheus does not rely on cloud infrastructure, resulting in better privacy, faster response times, and zero dependency on network availability.

We are also proud of how we fused multiple data sources into a single operational picture. By combining depth estimation with thermal intensity, we transformed raw sensor data into actionable insights. Finally, developing a flashover risk engine that warns firefighters before conditions become critical was a major milestone for our team.

What we learned

Building Prometheus taught us that life-safety technology demands more than accuracy—it requires reliability, clarity, and trust. We learned how closely hardware and software must work together when failure is not an option.

We also deepened our understanding of edge AI optimization and the importance of presenting complex data in a form that can be acted upon in seconds. Most importantly, we learned that the best technology supports its users quietly, without distraction.

What's next for Prometheus?

Our next step is taking Prometheus out of the lab and into real-world environments. We plan to collaborate with local fire departments to test Beacon during live training exercises and gather direct feedback from firefighters.

These field tests will help us refine the hardware’s ergonomics, improve the accuracy of our risk assessments, and expand our datasets to better reflect real fire behavior. Our goal is to ensure Prometheus is not only impressive in demonstrations but dependable in the field.