Engineering expertise for AI
Real engineering work, made useful for AI.
Caledonia Labs licenses real engineering project data and builds verifiable engineering tasks and environments for AI evaluation and training.
01 · Project data
Engineering project data
An archive of more than 1,100 real engineering projects with native, editable drawings, designed by our founder over his career, with the complete project files in hand.
What a package can include
- Native DWG and Revit drawings, and PDF drawing sets
- Technical reports, design memos and specifications
- Calculation sheets and bills of quantities, in part of the packages
- Software calculation models, such as slope-stability models, in part of the packages
Coverage
- Surveying and georeferencing
- Environmental studies
- Land regularization
- Architecture
- Land subdivision
- Water, drainage and sewer networks
- Earthworks
- Geotechnical reports
- Structures
A separate geospatial lot holds drone photogrammetry products (orthomosaics, point clouds and elevation models) and GNSS survey data.
We package projects around what you need, with drawings, calculations and reports kept together, and work by licence, sale or revenue share.
Everything is sanitized and anonymized under NDA before delivery: names, logos, signatures, registration numbers and file metadata are removed. Military and government-restricted material is excluded.
02 · AI tasks
Engineering tasks and environments for AI
Verified agentic tasks and RL environments for frontier-model evaluation and training, authored and reviewed by a civil engineer.
Each task ships with
Disciplines
- Civil
- Structural
- Geotechnical
- Hydraulic
- MEP
Also available
- Evaluation sets with verified answers
- Engineer review of model outputs
- Hard test cases drawn from real design decisions
We deliver in batches, starting with a paid pilot of 3 to 5 tasks. Our open RL environment built from a real geotechnical decision, geotech-rl, is public on GitHub and archived on Zenodo.
03 · About
Who we are
Caledonia Labs is led by Arthur Cunha da Silva, a licensed civil and geotechnical engineer with seven years in the field: designing real projects, running public infrastructure budgets for the Brazilian Air Force, and now building the engineering tasks that put frontier AI to the test.
Air Force. Arthur joined the Brazilian Air Force ranked 3rd in Brazil in the civil engineer-officer exam and served as a First Lieutenant, heading the engineering budget section in Natal. He led the technical and cost review of infrastructure across northeastern Brazil, from hospitals and aviation facilities to radar, water and power, and managed a team of engineers.
Design practice. Before that, as design engineer and project manager at a structural and geotechnical design firm, he took real projects from survey to approval: reinforced-concrete structures, slope-stability and dam studies, retaining walls, land subdivisions, and water, sewer and drainage networks approved by utilities and the fire department. Our data archive comes from this work.
AI evaluation. Today he authors benchmark tasks and reinforcement-learning environments that test frontier models on civil, structural, geotechnical, mechanical, electrical and chemical engineering, and publishes open-source engineering software validated against commercial packages.
Education. Civil Engineering at the Federal University of Viçosa (UFV), one of Brazil's leading engineering schools, and a postgraduate specialisation in Geotechnical Engineering at PUC Minas. Licensed by CREA-RJ.
We are based in Nova Friburgo, Brazil, and work in English with full overlap with US business hours.
04 · Contact
Tell us what you need
Which disciplines, formats and project types matter to you, or what kind of tasks your models need. We will reply by email.
[email protected]Or find us on LinkedIn.