Roof Measurement Software

The resolution cliff nobody warns you about

Tracing runs against imagery sampled at a fixed scale. That works until the roof exceeds the radius the sampling can resolve, and then the failure is sudden rather than gradual: the trace collapses, bleeds into an adjacent structure, or returns an area confidently wrong by a quarter.

The reason this is dangerous rather than merely annoying is which roofs it hits. Small simple roofs trace fine. It is the large, multi-structure, high-value jobs that fall off the cliff — the ones where an estimator most wants the help and is least likely to re-measure by hand.

The fix is to scale sampling with the roof rather than holding it fixed, cap the vertex budget so an outline stays a roof and not a noisy polygon, and clip bleed into neighbouring structures explicitly. None of it is glamorous, and all of it is the difference between a tool reps use and one they stop trusting after the second bad number.

Verify against real addresses, not a demo

The single most useful thing we built alongside the tracer is a harness that runs it against a set of real job addresses and reports traced area against a reference mask, corner deviation in metres, and how far the trace overshoots the actual roof.

It exists because measurement changes are deceptive. An adjustment that visibly improves one house can introduce a systematic error on another — a gap-healing rule that looks sensible band-fills every edge and inflates area by a quarter across the whole set. That is invisible on a single example and obvious across seven.

So nothing ships without a before-and-after run. It has caught two changes that would have shipped as improvements and were regressions.

Automation a rep can overrule

A measurement tool that cannot be corrected by the person standing at the property will be abandoned the first time it is wrong, and it will be wrong sometimes. Aerial imagery ages, additions get built, tree cover obscures edges.

So the output is editable: a rep can retrace a plane by tapping its corners, or add a structure the imagery missed. The automation is there to remove ninety percent of the work, not to insist it is right when the person on site can see that it is not.