Turning a two-day electrical takeoff into a review pass
Device counts extracted from construction plan sets by combining classical computer vision with vision models, measured honestly per category and reviewed by a human before it becomes a bid.
The problem
Estimating capacity capped how much work the company could bid. Reading a full drawing set for device counts took a senior estimator days, and every set skipped was a job never quoted. Off-the-shelf takeoff tools handled clean symbols and fell apart on real construction documents.
What we did
- 01
Combined classical template matching with vision models rather than betting everything on one technique — each covers the other’s failure modes.
- 02
Ran multiple independent extraction passes and aggregated by median with majority vote, so a single bad read cannot set a quantity.
- 03
Added dedicated focus passes for the categories that under-performed, instead of accepting a mediocre average across all of them.
- 04
Clamped declared quantities from the legend as hard upper bounds — if the drawing says twelve, the system cannot report fourteen.
- 05
Routed each drawing type to its own extraction path, because a panel schedule and a floor plan are not the same problem.
- 06
Reported per-category accuracy openly in the interface, so estimators know exactly where to look hard and where to trust it.
The transferable lesson
Publishing an honest 0.81 built more trust than claiming 0.99 ever would have. Estimators do not need a system that is always right; they need to know exactly where it is weak so they can spend their review time there. Accuracy you can point at beats accuracy you assert.