Inspiration

FireProof started with access to our own building and a question: if a fire started here, where would people get stuck?

We walk through buildings every day, but knowing where the exits are doesn’t tell us how an evacuation would unfold. A narrow doorway, crowded stairwell, or smoke-filled corridor could change which routes people can use. We wanted to make those interactions visible and explore how small changes affect the outcome.

What it does

FireProof turns floor plans and building models into an interactive 3D evacuation simulator.

Users can place a fire, configure occupants and door conditions, and watch simulated people move toward exits as smoke spreads. The dashboard shows evacuation times, exit use, queues, and smoke exposure.

FireProof is an exploratory prototype with an illustrative fire-and-smoke model, not a compliance tool or a replacement for professional assessment.

How we built it

We built the frontend with React, TypeScript, and Three.js, using React Three Fiber for the interactive scene, Recharts for charts, JuPedSim for simulating occupant behavior, and Zust and for application state.

Our Python pipeline converts traced floor plans or IFC building models into shared geometry. JuPedSim handles crowd movement, while our simplified fire-and-smoke model affects visibility, walking speed, and route availability. FastAPI connects the interface to simulations and batch analysis. We integrated Gemini to propose editable room geometry from floor-plan images and generate explanations of computed results. Simulation code produces the numbers. We also built a static playback mode with precomputed runs so the demo works without a live backend.

Accomplishments that we're proud of

We connected building input, multi-floor crowd simulation, smoke modeling, and interactive analysis into one workflow. We’re particularly proud of moving beyond an animation to answer actionable questions: where do queues repeatedly form, which exits matter, and what changes improve the simulated outcomes? One revealing finding was that similar evacuation times can hide different smoke exposure. Comparing both helps tell a more complete story than a single “time to exit” number.

Challenges we faced

One issue was pathfinding for people throughout buildings, as we initially built a naive approach towards the closest exits that didn't consider fire location. We fixed this by modifying human behavior to notice and adapt towards obstacles and pursue a different route towards an exit. Another issue was AI hallucination, as our LLMs produced too many random features that were relevant but not helpful. For example, the navigation bar was cluttered with too many options so we had to evaluate what features were important and necessary for architects.

What we learned

We learned that geometry is part of the simulation: door openings, stair connections, and exit placement can substantially affect results. We also learned to question our metrics. Faster evacuation alone doesn’t capture every outcome, and an intervention can improve one measure while worsening another. Finally, we learned that communicating assumptions is part of building a useful simulator. A compelling visualization needs clear explanations of what the model can and cannot support.

What's next for FireProof

We also want to make building imports more reliable, expand accessibility-focused scenarios, and test whether bottleneck rankings remain consistent across different occupant behaviors and smoke assumptions. We can improve simulating physics by understanding things such as HVAC, open windows, and how pressure differences affect smoke movement. Our goal is to help people understand their buildings better and bring clearer questions to professional safety assessments.

Built With

Share this project:

Updates

Submission history