From Earthquake Rubble to Queryable Structure in 4 Minutes

In February 2023, a 7.8 magnitude earthquake struck southeastern Turkey. Within hours, our team was on the ground with nothing but phones and a prototype algorithm.

The First Test

The challenge was immediate: rescue teams needed to understand building collapse patterns to prioritize search areas, but existing assessment methods took days. We had minutes.

Using phone cameras, we captured video walkthroughs of collapsed structures and processed them on a laptop in the field. The first reconstruction took 4 minutes and 12 seconds.

What We Learned

The experience fundamentally shaped how we think about spatial intelligence:

Speed matters more than perfection. A 90% accurate model in 4 minutes is infinitely more useful than a 99% accurate model in 4 hours when people are trapped under rubble.

Multi-modal input is essential. In a disaster zone, you work with whatever sensors you have. Phone cameras, drone footage, satellite imagery — the system needs to handle all of it.

Queryability changes decision-making. Once we could ask "show me all structures with >30° lean angle within 500m," rescue coordinators could prioritize systematically instead of going building by building.

# This query changed how teams were dispatched
at_risk = scene.query(
  type="building",
  lean_angle_gt=30,
  within=scene.bounds(center=epicenter, radius=500)
)
# → 47 structures identified in 0.3 seconds

That 4 minutes — from phone video to queryable 3D — became the founding moment for Percept.

From Prototype to Platform

Everything we've built since has been driven by that experience. The accuracy targets, the sensor flexibility, the query system, the simulation engine — all of it traces back to what we needed in Turkey and didn't have.

Today, the same pipeline runs in under 2 minutes with higher accuracy. But the core insight hasn't changed: spatial intelligence needs to be fast, flexible, and queryable to be useful when it matters most.