Skip to content
Case 01 · 8 min read

Case 01 · iOS · LiDAR · Insurance restoration

Wave Scan — a water-damaged room, measured in forty seconds.

Role
Capture & floor-plan geometry · first commit, top contributor
Stack
iOS 17+ · Swift 6 · SwiftUI · ARKit · RoomPlan
Ships
TestFlight from a green main · field-tested on real properties
Result
~40 s / room

A water-damaged room, measured in about forty seconds — by an app that knows when not to trust a measurement.

~40 s1
per room with LiDAR
10 in to squared2
error before squaring
18 m3
IMU drift after 60 s without tracking
239 / 426
non-merge commits

01 — The problem

Estimates priced on rooms nobody measured

After a flood, a restoration contractor writes an insurance estimate: square feet of drywall to cut, flooring to lift, days of drying. Those quantities are priced on room dimensions — and too often the dimensions are guesses.

Wave Scan is the companion iPhone app to the Wave estimating platform. A technician walks the property, the phone measures each room, and the real dimensions, wall areas, doors, windows and ceiling heights replace the assumed ones. The line items, and the money, move with them.

02 — What I built

The capture engine and the floor plan

I made the first commit and became the codebase's top contributor. I own the capture domain — the session that turns aims and taps into geometry — and most of the 2D floor-plan renderer, plus the release pipeline.

The architecture keeps the hard parts testable: rendering and the offline upload queue build and test on a Mac in seconds, without the app or a device. Floor plans are snapshot-tested against the same JSON fixtures the web renderer uses, so the phone and the browser draw the same room.

  1. LiDAR rooms through Apple RoomPlan in about forty seconds
  2. Hand-traced rooms on phones without a depth sensor
  3. Offline queue with resumable uploads that survives sign-out and retries
  4. Doors, windows and passages placed on the right wall

03 — The hard part

Corners from two aims

Most points in a room are easy: cast a ray from the camera onto a known plane — the floor, a wall. A ceiling corner is circular. The ceiling has no plane until its height is measured, the height is read off a wall, and walls only exist once corners close a ring.

So I borrowed the surveyor's method. The technician aims at the corner from two different positions, and I solve for the point where the two rays pass closest. Scale comes from the distance they actually walked, measured by the phone's motion sensors — real metres, not camera guesses.

The solver rejects degenerate aims, too short a walk, too shallow an angle and points behind the camera. How far the two rays may miss each other grows with range — about 0.24 ft at 5 ft and 0.82 ft at 20 ft — because a fixed bound is wrong in both directions.

04 — Knowing when not to measure

Refusing the taps that lie

A ray aimed at a sofa passes straight through it to the floor, so the corner lands feet away from the wall. One early field trace put a wall through the middle of a sofa. Now the app clusters the scene's feature points along the aim and refuses taps that hit furniture or a soffit — and shows the technician why.

Without LiDAR, motion-sensor drift grows with the square of time: half a metre after ten seconds, eighteen metres after a minute. When ARKit loses track of the scene, nothing corrects it, so every tap is refused until tracking returns.

Each refusal is a choice: the furniture check fails open on sparse evidence, so bare floors don't nag; the tracking check fails closed, because a wrong room is worse than a slower one.

05 — Honest geometry

Squaring rooms traced by hand

Thumb tremor compounds around a room. The first hand-traced room closed with opposite walls ten inches apart. Nudging corners fights itself, because every corner belongs to two walls — so I snap wall directions to right angles relative to the longest, most accurately measured wall, then re-intersect.

Every ceiling height carries its provenance: assumed, stated or measured off a wall. The highest consistent ceiling wins over time, a wall-aimed reading is never overruled, and an assumed ceiling reaches the estimate flagged as an assumption.

06 — What I'd change

Next

A tape-measure accuracy study to replace the provisional squaring tolerance with a measured one, and whole-house walks that stitch rooms beyond RoomPlan's single-session limits.

Notes

  1. Approximate capture time per room on LiDAR iPhones.
  2. Opposite walls of the first hand-traced room closed 10 inches apart over about 16 ft.
  3. A 0.01 m/s² bias grows as time squared: 0.5 m in 10 s, 18 m in 60 s.

Ask my portfolio