Thomas Czerniawski is the lead machine learning engineer at Integrated Projects, where he is automating scan-to-BIM, the conversion of laser scans of existing buildings into usable 3D models. He trained as an engineer at the University of Waterloo and did graduate research in reality capture and computer vision. He has spent his career turning information that exists but cannot be used into something a decision maker can read, which makes him a clear guide to what automation can and cannot do for buildings.

Thomas Czerniawski, Lead ML Engineer, Integrated Projects

Watch the episode ↗

In this conversation

The week-long spreadsheet that showed a $20 million surveying quote on a Toronto subway project was about twice what the work required.

Thomas on laser scanning as a bridge between the physical and digital worlds, from pipe spools to a hotel foundation nobody had drawings for.

His case that the technology that will make 1.6 billion buildings useful is the one that hides almost everything we capture.

Why a library of thousands of known modeling errors makes scan-to-BIM automation both hard and possible.

His answer to whether AI will take the BIM specialist's job, told through vitamins, Edison's light bulb and a grove of Oxford oaks.

Scan-to-BIM is a technology that holds the promise of shortening the time between a question and an answer.

Thomas Czerniawski

Chapters

The spreadsheet that halved a quote

A Waterloo co-op student turns a tangle of project documents into one readable table.

Laser scanning as a bridge

From a briefcase-sized battery to scanners that are phones, and what a digital replica lets you do.

Hide most of it

Why making information useful means choosing what to leave out.

From question to answer

A century-old hotel, missing drawings and one sentence in a 1931 journal.

By the numbers

The Claridge's basement extension, London

new basement levels dug under a working hotel

existing building columns carried on new foundations

30 m

depth of hand-dug shafts under each column

5,500 m²

of space added

Thomas uses this project to show the cost of not knowing what is inside a building: the engineers had to find the foundation's steel in an archive before they could dig at all.

Source: Arup, Claridge's basement extension ↗

Why automation is hard, and the roadmap

Thousands of QA issues, a floor-area patent, and walls, doors and windows next.

What people do when the drudgery is gone

The hierarchy of work, and a lesson in stewardship for young engineers.

In automation, variability is the enemy.

Thomas Czerniawski

Takeaways

Most expensive guesses in construction come from information that exists but is not accessible. Combining and filtering it into one simple view, even in a spreadsheet, can cut a cost estimate in half.

Automation fails when its output takes longer to fix than modeling from scratch. A standardized data set and a tight loop with the people who review the models matter more than the algorithm.

Start with the questions everyone asks, such as floor area or how many windows there are, and deliver them as a menu of use cases rather than a model people have to learn.

We want to make building information so cheap that only the rich will be able to go out on site and do a physical inspection.

Thomas Czerniawski