Skip to content
Kamai
Geometric foundational models

Geometric foundational models,built for construction.

General-purpose AI looks at a drawing and sees a picture. Kamai builds vertical AI for construction: foundational models developed and trained in-house that read drawings the way an estimator does - geometry, scale, symbols, and intent.

See it on your sheetsBuild on the API
Wall detectionScale resolutionSymbol recognitionDimension parsingCross-sheet referencesHatch classificationLayer semanticsGPU renderingExact measurementSchedule extraction

The model stack

Not one model. A foundational system.

Kamai is a family of special-purpose models we develop and train in-house, composed into a foundational system built specifically for the problem of reading construction drawings. No single model can do this - the system is the breakthrough.

01

Purpose-built AI models

A stack of models we train and develop ourselves for drawing understanding - not a general-purpose vision model bolted onto blueprints.

02

Geometry as the native input

The models read coordinates, scale, layers, and symbols straight from the file, so every measurement is exact and traceable back to the source.

03

Composed into one system

The models work as a single foundational system, generalizing across disciplines and sheet types without per-project re-training.

Vertical AI

Depth in one domain beats breadth across all of them.

Horizontal models are trained on everything and master nothing. Kamai goes vertical: every model in the stack exists for one job - understanding construction drawings.

General-purpose AI

  • Rasterizes the sheet and reasons over pixels.
  • Trained on the whole internet; fine-tuned on nothing in particular.
  • Estimates quantities it has no way to verify.
  • Needs per-project prompting and tuning to stay usable.

Kamai vertical AI

  • Reads the geometry the drafter actually drew.
  • Special-purpose models trained in-house on construction drawings, symbols, and sheet conventions.
  • Every measurement computed from the drawing's native geometry - not estimated.
  • Generalizes across disciplines and sheet types without re-training.
The runtime underneath

Foundational models need a foundation.

The models run on a GPU-accelerated geometric runtime built for construction-scale drawings. These four ideas are it.

Native drawing ingestion

Parses PDF primitives directly. No rasterization. No OCR fallbacks except where intentional. Kamai reads what was drawn, not what was rendered.

GPU-accelerated rendering

The viewer draws vector geometry on the GPU rather than serving image tiles. Pan and zoom stay smooth on full sheet sets, and linework stays sharp at any zoom depth.

Semantic geometry layer

Every shape is typed: wall, door, dimension line, hatch, symbol. The models reason over structured geometry - scale, layers, and cross-references intact.

Exact, auditable measurement

Lengths, areas, and counts come from the drawing's native coordinates, and every number traces back to the geometry that produced it.

From geometry to decisions.

The geometric runtime surfaces what is on the sheet. The foundational models explain what it means.

For builders

Every primitive Kamai sees is available through the API.

Build on the same models our customers use. Self-service API access is live - sign in to the Kamai Console to get started.

Open the ConsoleExplore the developer hub
FAQ

Technology questions, answered.

A foundational model built for geometry rather than language or photos. Kamai trains special-purpose models in-house and composes them into one system that reads construction drawings - geometry, scale, symbols, and cross-references - and generalizes across disciplines and sheet types without per-project training.
General-purpose models treat a drawing as an image and estimate. Construction needs exact quantities, so Kamai goes vertical: foundational models trained specifically on construction drawings that compute - not guess - every measurement. Depth in one domain beats breadth across all of them.
No. Most AI takeoff tools convert blueprints into images, then run computer vision over the pixels. Kamai works directly on the geometry stored inside the PDF or CAD file. Shapes, scale, and text remain machine-readable at every step. No information is lost to rasterization.
Digitally drafted PDF (including layered PDFs). Raster PDFs and scanned drawings are supported with a documented OCR path, with clear flags so you know when you are outside the native vector pipeline.
Measurements are computed directly from the drawing's native coordinates, so dimensional accuracy is bounded by the precision of the source drawing - typically sub-millimeter. Kamai never approximates a measurement from a pixel grid.
Yes. The renderer is GPU-accelerated and draws vector geometry directly, so multi-thousand-page sets stay responsive. Cross-sheet references resolve in the same session.
Yes, on-prem and VPC deployments are available for enterprise and regulated customers. Contact us to scope a deployment.
Keep reading

Guides on AI takeoff.

See for yourself

Bring a sheet. See what Kamai sees.

Try it nowTalk to sales