EvacFlow

Inspiration

Emergency evacuation plans are often static: a floor map is posted on a wall, but it does not show how congestion, blocked paths, or changing conditions can affect evacuation.

We wanted to explore whether a school could turn its existing floor plans and evacuation maps into a more interactive planning tool. That idea became EvacFlow — a platform that helps administrators model evacuation routes, identify possible bottlenecks, review physical constraints, and prepare guidance for drills and emergency situations.

What it does

EvacFlow converts building and floor-plan information into a structured evacuation model.

Administrators can:

• import and review floor plans or evacuation maps
• define rooms, hallways, stairs, and exits
• calibrate physical dimensions such as corridor and exit widths
• simulate how groups move through the building
• compare route and exit utilization
• identify queues, congestion, and potential bottlenecks
• review and publish evacuation guidance for participants
• manage evacuation drills and operational sessions

EvacFlow separates simulation results from published guidance. A simulated route is treated as planning information and must be reviewed before it is presented to participants.

How we built it

We built EvacFlow as a web-based planning and simulation system with a structured building model at its core.

The routing layer represents rooms, hallways, stairs, and exits as a graph and calculates available evacuation paths. A physical simulation layer adds real-world concepts such as distance, walking time, corridor capacity, throughput, queue delay, and congestion.

We also developed a local AI-assisted floor-plan import workflow using Ollama and Qwen3-VL. Instead of automatically trusting AI output, detected geometry and labels are treated as reviewable proposals. Ambiguous measurements and connections remain unresolved until an administrator confirms them.

For real-world testing, we used photographed school evacuation maps rather than clean architectural CAD files. This exposed several practical challenges around vision-model reliability, memory usage, structured output, and geometry extraction.

Challenges we ran into

One of the biggest challenges was translating a visual floor plan into reliable topology.

A vision model may successfully recognize rooms and labels while still producing unstable hallway geometry or duplicate objects. We found that asking the model to extract too many categories at once could cause extremely long or truncated outputs.

To address this, we experimented with smaller, isolated extraction stages for hallways, stairs, exits, and room labels, combined with strict validation and deterministic processing in application code.

Another challenge was local AI resource usage. Large context windows caused significant memory pressure, so we added bounded context sizes, explicit timeouts, and model unloading to keep local inference predictable.

We also had to be careful not to turn uncertain AI detections into false certainty. EvacFlow therefore preserves review-required and unresolved states instead of guessing missing connections or physical dimensions.

What we learned

We learned that emergency-planning software needs more than a shortest-path algorithm.

A useful evacuation model must consider:

• multiple exits
• shared corridor congestion
• stair capacity
• queues
• physical dimensions
• blocked facilities
• uncertainty in imported building data
• human review before operational guidance is published

We also learned that AI is much more useful as an assistive extraction tool than as the final authority. Deterministic validation, provenance, and administrator review are essential when model output affects safety-related planning.

What's next for EvacFlow

Next, we want to improve real-world floor-plan extraction, validate the system with additional school layouts, and make calibration easier using measured building data.

We also plan to continue improving evacuation simulations, operational drill workflows, participant guidance, and location-aware assistance while keeping administrators in control of what guidance is actually published.

Built With

• accessibility
• algorithms
• cloudflare
• computer
• d1
• dijkstra
• events
• geospatial
• graph
• next.js
• node.js
• ollama
• pathfinding
• pwa
• qwen3-vl
• react
• server-sent
• simulation
• sqlite
• three.js
• typescript
• vision
• vite
• webgl