Estimators spend days counting symbols, walls and facade areas across hundreds of plan sheets. I build the vision models and the agentic layer that do that reading for them, inside Boon's takeoff product.
One-shot symbol detection
Point at one symbol in the drawing legend and the model finds every instance across the plan set. Open-vocabulary, so a symbol it never saw in training needs no new labels and no retraining.
Facade segmentation
A segmentation model built from the ground up for building elevations, so cladding, glazing and base areas come out of a drawing as measured regions instead of hand-traced polygons.
Agentic association
Detections mean little on their own. A VLM-driven agent ties each one to its legend entry, schedule row and spec note, checks the result and keeps a trace of why.
- Re-engineered a production wall-takeoff model and its inference path: 2.6× faster, 40 to 60% less GPU memory, about 5% more accurate.
- Electrical takeoff and branch-circuit inference on dense plan sets.
- A real-time voice agent that calls drivers and turns the call into structured operational data.
10×
0.90+
2.6×
SentBe →