Research
A spatial foundation model for construction drawings.
Most construction AI is a language model reading text pulled out of a PDF. We made a different bet: a takeoff is a measurement of physical space, so the model has to see geometry — boundaries, scale, and the relationships between elements across a whole set. One model body trained across every trade, with specialized heads on top. Below is the published work behind that bet, and the measured results it produces, including where it fails.
The architecture bet: geometry, not language
A takeoff is a measurement of physical space. Text extraction throws away the thing you are trying to measure.
-
Why Estimating Is a Geometry Problem, Not a Language Problem
The industry is reaching for LLMs to read drawings. That solves the wrong problem — quantities live in geometry, scale, and spatial relationships, none of which survive flattening a sheet into text.
-
One body, many heads: the spatial foundation model behind our AI construction estimating
Most construction AI is a drawer full of single-purpose models. One foundation model body trained across every trade, with specialized heads on top, transfers what it learns between trades instead of hitting a per-detector ceiling.
-
On reading construction drawings
A third kind of reading the AI industry has not built yet — neither natural-language reading nor photographic recognition — and the field that needs it most.
-
The 400-page drawing set is a database in disguise
Spec, plan, and schedule disagree, and an RFI is filed weeks later. The set is a spatial database with no query layer, which is what makes cross-sheet reasoning the real product.
-
Beyond Construction: Why a Spatial Foundation Model Eats the Whole AEC Stack
A general contractor and a facilities manager have almost nothing in common until you hand them the same drawings. The spatial layer generalizes past preconstruction.
-
The Vertical AI Doctrine: Why Horizontal Wrappers Can't Win Construction
Horizontal AI wins legal research and support. Three structural properties of construction documents stop a general-purpose wrapper from winning here.
How the model is built and trained
Agentic vision instead of single-pass recognition, and a synthetic data factory to cover a distribution real-world collection cannot.
-
From Pixels to Decisions: Building Agentic Vision for Construction Diagrams
Isolate relevance, crop intentionally, decompose the task, parallelize reasoning. The difference between recognition and reasoned perception on sheets that are 70–80% visual noise.
-
How We Are Solving the Training Data Problem for Construction AI
Thousands of synthetic edge-case blueprints per hour — degraded scans, overlapping annotations, non-standard legends — to fill the gaps real drawing collection never reaches fast enough.
-
Why MEP coordination is actually a multi-page spatial reasoning problem
A panel schedule on one sheet, the equipment it feeds on another, the riser diagram on a third. Coordination requires holding a set in mind, not a page.
-
The Data Flywheel: How Proprietary Construction Vision Creates an Unreplicable Moat
Why the company that builds the best construction-specific training data compounds fastest over the next decade.
Measured results and failure modes
Per-trade F1 against human-annotated ground truth, published including the trades where we are weakest.
-
The Trust Chasm: Why Construction AI Accuracy Is the Only Metric That Matters
The full per-trade F1 table from MLflow evaluation runs: steel 0.93–0.95, HVAC fittings 0.84, ceiling segmentation 0.997, electrical devices 0.64 — plus the three-stage path that moved most trades 15–25 points in six months.
-
If a Vendor Won't Talk About F1, Walk Away
The one question that separates a measured system from a marketed one, and the accuracy floors we publish: electrical ≥ 0.90, architectural ≥ 0.85, mechanical ≥ 0.74.
-
How wall-type takeoff reaches 82% accuracy
Detection is close to solved; association is the real work. Mask F1 0.842 and boundary F1 0.92, but end-to-end ~0.72 because the reader sometimes locks onto the wrong schedule on multi-schedule sheets.
-
How Boon AI Detects Mechanical Ductwork with 85% Accuracy
Every straight duct segment — horizontal, vertical, diagonal — with width measurements, and what the remaining 15% actually consists of.
-
Why 78% accurate feels wrong — and why that's the wrong question
A hospital set came back at 78% quantity match, and the estimator still had a reviewed, bid-ready number in 22 minutes. Time to reviewed takeoff is the metric that maps to margin.
The short version
- Takeoff is geometry. Text extraction discards the scale and spatial relationships that produce the quantity.
- One foundation model across trades beats a drawer of single-purpose detectors, because representations transfer between trades.
- Agentic vision — decide what to look at, crop, decompose, parallelize — beats single-pass recognition on sheets that are 70–80% noise.
- Synthetic data is the unlock for a drawing distribution too wide to collect through.
- Accuracy gets published per trade, weakest trades included, or it isn’t a measurement. See the table.