Skip to content

Machine learning: investment of a ton yields millions annually

In 2019, a supplier of air conditioning systems with software technology company Luminis is sharing a potential market opportunity that he has been racking his brain about for months. His company is active in the office market and supplies large buildings with ventilation systems. Now he also sees potential in smaller projects, especially in the more luxurious private sector. There is demand there, but it just doesn't seem to work out a viable business case.
Climate Control


This is the headline

One-time investment in machine learning earns this business case millions annually

This is the headline

Machine learning: extra capacity with the same people

This is the headline

Use of open source algorithms accelerates time to market

This is the headline

Automation enables expansion into markets that were previously inaccessible or non-lucrative

Gerard Verbeek, director of Luminis in Apeldoorn, listens to the story. His new relationship develops and makes indoor climate equipment and does the planning around it. In order to integrate the systems efficiently, his employees have to delve into construction drawings. Verbeek: "They add circulation ducts and air renewal equipment to existing construction drawings and calculate the supply and return of air for all rooms based on content and building regulations." This is primarily manual work. Individuals have a day's work studying and calculating drawings and fitting valves, ducts, airlocks and equipment. In the market for large projects and offices, these labor-intensive operations can be justified, but in the private market it is hardly worth the effort.

Scale up or automate?

Scaling up in this market is an option, but that again brings overhead. "Added up, it's relatively a lot of work relative to the benefits," says Verbeek. "The client actually wanted the number of employees for the planning work to remain the same. So that was crying out for automation." Verbeek points his client to the existence of open source algorithms that can help recognize drawings. Based on this, it is possible to build an application suitable for the private segment the installer wants to target. This eventually resulted in a web application in which the architect or contractor uploads drawings. These are often pdf's, but a scan or photo of a physical drawing is also possible. With this approach via a web interface part of the manual work shifts to the end customer.

"The customer specifically wanted the number of employees for the planning work to remain the same. So that was crying out for automation."

Classification through machine learning algorithms

In a next step, the system classifies the drawings with machine learning algorithms. "The function for recognizing the drawing is open source and we didn't have to program anything else for it," Verbeek clarified. Doors, walls, but also kitchens, living rooms and other spaces are recognized by the algorithms with approximately 80 percent certainty. The latter is characteristic of machine learning. The algorithms are trained and capable of recognizing objects or drawings, but the judgment is never 100 percent watertight. There is always a degree of uncertainty in the outcome.

So the system makes mistakes, but the machine learning algorithms can be improved, making the classification better and better. "Employees are shown an outcome and degree of certainty. For example, 'a living room with a 70 percent chance that this is indeed a living room.' If they confirm that, then the algorithm learns from that." Because the system provides the recognition with a statement about the probability, users are able to choose the settings in such a way that they only have to look at the real doubtful cases. "For example, to all the statements made with a certainty of below 90 percent."

"Employees are shown an outcome and degree of certainty. For example, 'a living room with a 70 percent chance that this is indeed a living room.' If they confirm that, then the algorithm learns from that."

AutoCAD as the next step

Meanwhile, it is also possible to import original AutoCAD files. AutoCAD is a drawing tool that construction uses a lot. Digital AutoCAD files have the advantage that they contain metadata. As a result, some of the parts can be classified with 100 percent certainty. "Our final product is still an enriched PDF," says Verbeek, "but a translation to AutoCAD is among the future possibilities." At the start, the customer must feed the system with labeled data to train the machine learning algorithms. "When applied in practice, more information is added with each upload and the system learns as employees check and agree on the labels," Verbeek says. "You can quickly add value and fine-tune." After interpretation, employees enrich the drawings with existing drawing tools. "In the beginning it's a bit more intensive, more work, but that gets less and less, because the system gets better," Verbeek says. "Very occasionally someone still has to look at it, but the errors are decreasing, because they are gradually training the system."

Turnaround time of nine months

The total project has a lead time of nine months. By early 2021, the result will be a closing, profitable business in an adjacent market. "The customer expects to serve a market that is potentially 10 million euros in size. With an estimated investment of 100 thousand euros, the expected payback period is less than a year, if you assume that the application saves two full-time workers. One of the additional benefits is that the customer also expects to be able to apply it in its existing market, the professional segment," Verbeek said.


This is the headline

Scaling up is an interesting option but may involve overhead. In the business case, consider the option to automate for valuable time and/or cost savings.

This is the headline

Explore the potential for using open source algorithms. This can be the basis for building an application that supports the ultimate goal of the business case.



Luminis is a group of companies, headquartered in the Netherlands, that specializes in providing innovative solutions to business and government, primarily using emerging (information) technology.

View Business
Sign up for our newsletter and receive updates.