On a foundry production line, two operators are inspecting the same metal part. Both identify a defect. Yet their conclusions differ. One believes the part can continue through the line, while the other believes it must be rejected. This situation is far from unusual in the industry.
Behind every quality control check lies a degree of human expertise—often intuitive—that is difficult to formalize and even more complex to convey to a machine. Contrary to some common misconceptions, artificial intelligence does not become effective simply by being fed thousands of images. It is also necessary to know what information to include, what to discard, and, above all, to understand what experts are actually looking for.
It is precisely at this intersection of industry, technology, and human understanding that Gabrielle Van de Vijver, a vision application engineer at Psycle, comes into play. Having joined the company nearly two years ago after a post-graduation internship, she is now involved in developing machine vision systems capable of assisting operators without ever replacing their expertise.
Is the operator the best teacher for the algorithm?
Recently, Gabrielle worked on a project related to the detection of foundry defects. To understand the client’s expectations, she spent nearly a full week working alongside the operators. She observed the production lines, engaged in informal discussions, analyzed parts, and gained an understanding of the defects. The teams even went so far as to share with her old technical manuals dating back to the 1980s.
“In a conversation with an operator, I learned that a visually obvious defect is corrected by light sanding. Conversely, a detail imperceptible to the naked eye compromises the very structure of the part. I aim to identify these types of issues before feeding them into the algorithm.”
The challenge quickly becomes apparent. Not all operators use exactly the same criteria. Practices differ. So do interpretations. It therefore becomes necessary to gradually develop a common language—frameworks that are understandable to both domain experts and artificial intelligence. For Psycle, the goal is not to make AI omniscient, but rather to push it to refine itself so that it becomes useful.
“The goal isn’t to bring together all the expertise in the world into a single artificial intelligence. What matters is that it be reliable when it comes to the problem we ask it to solve.”
Understand the business before automating
Psycle’s projects regularly lead teams to explore fields they are unfamiliar with: foundries, the food industry, radiography, biology, aerospace, and logistics. For Gabrielle, this diversity is one of the most rewarding aspects of the job.
Certain biological categories, for example, can manifest themselves in several different visual forms. Whereas a mechanical part generally meets relatively stable criteria, living organisms introduce a much greater degree of variability.
“I’m fascinated by biological subjects. Living things don’t follow a set of manufacturing instructions; they defy strict categorization. Algorithms must learn to handle ambiguity and unpredictable phenomena.”
A vision project isn’t limited to just a camera
When people hear the term “machine vision,” many immediately picture a camera monitoring products on a conveyor belt. But the reality is much more complex.
Behind every system lie dozens of parameters that must be controlled: lighting, object movement, depth of field, mechanical constraints, production rates, communication with PLCs, and the interfaces used by operators.
This complexity explains why projects led by Psycle regularly draw on several complementary areas of expertise:
“It’s easy to identify the author of a set of specifications or a code. Each team member has a strong expertise, whether in optics, mechanics, or computer science. Psycle strongly encourages this complementarity.”
This approach has also supported the company’s recent growth. Since she joined, Gabrielle has seen the workforce grow significantly and projects take on greater scope. The offices have changed, and so have the teams. The versatility of the early days is gradually giving way to increasingly specialized expertise.
A blend of technology and pedagogy
One of the ideas that comes up most often during the interview is communication. Contrary to popular belief, the challenges aren’t always technical. In fact, the main obstacles are sometimes human.
“The sheer volume of data can be overwhelming. Experts in the field are often taken aback at first.”
His role, then, is to make the subject concrete. To show an image. To illustrate a flaw. To start from the craft itself. Because essential information is almost never provided spontaneously. So you often have to go out and find it and know how to ask the right questions.
Gabrielle owes this approach in part to her education. After completing a science program that combined mathematics, artificial intelligence, and data science, she chose the “Humanities and Technology” track at UTC, convinced that an engineer is never content to simply follow instructions.
“We also have a responsibility to understand what our choices entail and what effects they have on those around us.”
This conviction shapes her entire approach to work: developing an AI system isn’t just about striving for the best performance, but about understanding what we’re really trying to measure—and, above all, why. She has little interest in omniscient models. For her, technology must remain a tool in the service of human expertise—not the other way around.
Safety Shoes: A Wise Choice
There’s one last story to tell: the story of how she joined Psycle.
At the time, Gabrielle was finishing her studies and had a pair of safety shoes she no longer needed. So she posted an ad on a UTC Facebook group to sell them. Days went by without a response. Then a message arrived:
At the time, Gabrielle was finishing her studies and had a pair of safety shoes she no longer needed. So she posted an ad on a UTC Facebook group to sell them. Days went by without a response. Then a message arrived:
“I’m a data science engineer, too! But I wear my safety shoes every day. You should keep them just in case.”
The conversation ended there, however, since the shoe size didn’t match. A few weeks later, Gabrielle came across a particularly interesting internship opportunity. When she looked at the name of the person who posted it, she immediately recognized the young woman. Today, they work together at Psycle. And the shoes are still there.
“In the end, they were never sold. I use them almost every day now. I promised myself that if a computer science student ever wanted to sell hers, I’d tell her exactly the same thing.”
And in the end, we realize that Gabrielle is her own artificial intelligence—an intelligence that draws on customer knowledge, their lines of business, and their constraints.
And she always wears safety shoes (dust-free).