Thursday, 9 April 2026
On a production line, everything moves fast. Very fast. The specifications Psycle receives indicate speeds sometimes exceeding 300 products per minute. At this pace, an error is not always visible to the naked eye, but it can be costly: line stoppages, material losses, non-conformities, even product recalls. In […]
On a production line, everything moves fast. Very fast. The specifications Psycle receives indicate speeds sometimes exceeding 300 products per minute. At this pace, an error is not always visible to the naked eye, but it can be costly: line stoppages, material losses, non-conformities, even product recalls.
In this context, machine vision is establishing itself as a key tool. But between the technological promise and the reality in the field, a gap remains. Because for an artificial intelligence model to work, everything around it must be reliable. This is precisely the part that Gaëtan Blond, data engineer at Psycle, works on.
Having joined nearly five years ago after a final-year internship at UTC, he now works at the heart of the company's technical architecture. His job is to make possible what would otherwise remain theoretical.
Architecture first and foremost
Before an algorithm makes a decision, the information it receives must be structured. On a line equipped by Psycle, cameras continuously capture images. This data must be processed without latency or loss, then turned into decisions: compliant or not, part accepted or ejected.
Between these steps, a complex technical chain comes into play. Gaëtan designs the tools that collect these streams and, above all, make them usable.
But in some cases, several thousand images are generated every minute. This work therefore requires careful resource management. You have to make trade-offs, prioritize and organize data loading and unloading cycles to keep the system stable. Because AI does not fix a poorly structured system; it depends on it.
Making AI work
Once in production, the artificial intelligence model analyzes the images and produces a result. This result is then used to make an immediate decision on the line.
“In industry, this step tolerates no approximation. A false negative can let a critical defect through. A false positive can slow down production unnecessarily. So it's a constant balancing act.”
Gaëtan acts as a facilitator here. He does not develop the models directly, but he builds the environment in which they operate. His work thus determines Psycle's ability to adapt quickly without compromising the reliability of the systems in place.
Retraining without disrupting production
A model is never frozen. Products evolve, defects change, production conditions vary. It then becomes necessary to retrain the algorithms. This work makes it possible to improve performance without interrupting production.
Anticipation as the central point
According to some estimates, one hour of production downtime can cost between €10,000 and €100,000 depending on the sector.
To avoid these situations, Psycle implements continuous monitoring of the machines deployed. Each computer reports indicators in real time.
Gaëtan helped structure this system. Data is centralized on secure servers, analyzed, then turned into alerts when certain thresholds are exceeded. The team can then intervene remotely, often before the customer even notices a problem.
The reality in the field
In the field, conditions are rarely ideal. Some installations take place in sensitive areas, with strict access constraints. Others involve older equipment that has to be integrated without being replaced.
These situations demand constant adaptability. They are also a reminder that technology is only worth its ability to work in the real world. And Psycle's teams are well aware of that.
Security and industrial standards
Some projects take place in environments subject to strict standards, such as ISO 27001. This requires rigorous practices in terms of security and information protection.
Gaëtan contributes to this dimension by setting up secure connections, data encryption and flow control.
Industrial AI is not just about performance. It must also operate within a secure, controlled framework.
An eye on the future
Since his early days at Psycle, Gaëtan has seen the capabilities of artificial intelligence evolve. Hardware constraints have changed, tools have improved and possibilities have expanded.
Before integrating a new technology, tests are carried out to measure its relevance: performance, cost, ease of integration. This technology watch helps guide technical choices and offer viable solutions to manufacturers.
A passion that goes beyond the screen
Behind this career lies a long-standing curiosity for computer science. Developed very early, pursued as a self-taught interest and then deepened at the University of Technology of Compiègne (UTC), it now goes hand in hand with his involvement in the France-ioi association, where Gaëtan helps young people learn algorithmics.
Today, he remains committed to passing on knowledge, particularly to young audiences. Outside of work, he also devotes time to the SPA de Compiègne animal shelter. A necessary balance, away from screens, that allows him to keep a concrete connection with nature.
Artificial intelligence does not rely solely on algorithms, but on women and men capable of making it reliable and useful.
And as we know, in industry, it is often what you don't see that makes all the difference.