Overview
Open source

Open data, tools and models for construction

What helps the whole industry, we publish: reference data, prompts, agent skills, schemas and two computer-vision models. The open models are lite versions — clients get the full, latest models, currently in private beta.

Join the private beta Request demo access

What is published

Repositories and models

Everything below is free to use under its own licence. Issues and pull requests are welcome on GitHub — including on ConstructBench, our open benchmark for construction AI.

GitHub8 repositories

Hugging Face2 models

On GitHub constructelligence-lab

On Hugging Face constructelligence

Construction site safety hazardsconstructelligence/construction-site-safety-hazards

Detects workers, hard hats and hi-vis vests — worn and missing — and heavy plant (excavators, loaders, dump trucks) in site photos and video. Runs offline in a browser tab.

Object detectionYOLOv8n · 12 MB ONNXAGPL-3.0Electrical circuit connectivityconstructelligence/electrical-circuit-connectivity

Reads an electrical floor plan and returns which devices sit on which circuit, the runs between them and which circuits have a home run.

Floor-plan readingU-Net + graph decoder · ONNXApache-2.0

Both models ship as ONNX with the decode code, so they run offline — in Python or straight in a browser.

ConstructBench is an AEC benchmark built around construction and real-world applications — estimating, takeoff, scheduling and cost work, scored the way a contractor would judge it. We are releasing our own models' results against the leading general-purpose models as they land.

Our own benchmark

ConstructBench

An AEC benchmark focused on construction and real-world applications. Every task is drawn from the documents, drawings and decisions a contractor actually works with — an estimate, a schedule, a contract, a month of job cost — and scored the way a contractor would judge the answer, so a model that is genuinely good at construction shows it and one that only sounds confident does not.

The forecasting models behind the product are trained on proprietary construction cost data and stay private. ConstructBench is the public side of that work: an open harness, openly scored, with our models' results against the leading general-purpose models released as each run lands.

GitHubconstructelligence-lab

Hugging Faceconstructelligence

What it measures

Estimating & takeoffQuantities read from a drawing set and priced against a cost code, within the tolerance a real bid can live with.

SchedulingDurations and sequence from a scope and a crew, and the finish-date risk that follows from them.

Cost & forecastingCost at completion from a job’s own burn rate — the number the whole platform turns on.

DocumentsContracts, specifications and RFIs: the notice period, the exclusions, the clause that does not flow down.

Field & progressWhat is actually installed, read from logs, photos and quantities rather than from what was claimed.

Reference dataCost codes, units and waste factors — the unglamorous data every other answer depends on.

How it is scored

Real job material. The sample set a contractor would recognise, not puzzles written for a model.
A contractor’s tolerance. A quantity or a cost is right, within a stated band, or wrong — partial credit is explicit, never hidden in an average.
Per task and overall. A model that is strong at takeoff and weak at forecasting is visible as exactly that.

Results — our models against the leading general-purpose models, published here and on Hugging Face as each run lands. See the ConstructBench leaderboard — coming soon →

More of the platform

Your numbers are already there. Go look at them properly.

Hosted for you or self-hosted, read-only, every user included. Connected and reconciled against your own reports in a single 90-minute session.

Join the private beta Request demo access