Beta
Where this stands — measured, not claimed
Most software in this trade tells you a feature works. Here are the actual numbers off our production telemetry, including the parts that aren’t finished yet.
89%
of engine-detected counters were kept with no shape edit at all — no vertex drag, no merge, no re-dimension
73 → 97%
first-shot clean rate across the session history — the first thirty detections versus the most recent thirty
70%
of pieces came from the engine rather than being traced by hand
34%
of the remaining human corrections are edge typing, not shape fixing
Measured across 40 takeoff sessions, 364 sheets classified and 129 committed regions. Per document the clean rate ran from 100% and 96% on vector plan sets down to 33% on a scanned one.
The honest caveat. If somebody deletes a detected piece and traces it themselves, that leaves no correction behind to count — so the true “engine got it right” rate sits somewhere between 62% and 89%, not at 89% flat. We’d rather publish the range than the flattering end of it.
And what’s still manual. Shape detection is largely solved on vector plan sets. Edge typing is not — it’s the single biggest remaining click cost, and it’s what we’re working on next. Scanned sheets are the weak spot too, though they were only 3.4% of the thousand-page corpus we measured.
Every correction anyone makes is stored against the raw detection, so the engine learns the difference. That’s why the number moved from 73 to 97 — and why we want your plan sets. Book a demo and bring one, we’ll run it live and you can watch where it wins and where it doesn’t.