The customer is a Flemish manufacturer of micro-ingredients and functional additives for the animal feed sector, with customers worldwide. Its main site in Belgium brings together production, distribution, the lab and commercial operations, employing more than eighty people. An ISO-certified environment with strong quality discipline, with no previous AI applications on the shop floor.

Flaws in the seam finish of filled bags are barely visible to the naked eye. They only come to light when a bag leaks during transport, resulting in rejected containers in export countries, financial loss and reputational damage with international customers.

Several lines run continuously, so checking every bag by eye isn't feasible. A sample check will, by definition, let defects through. The question was clear: inspect every bag without slowing down the line.

The project started broadly across the whole site. Production staff, quality inspectors, logistics personnel, the lab and management each shared their pain points, at their own pace, through an AI bot. Visual quality control of bags after filling emerged as the priority.

In a half-day workshop, the use case was weighed up. High on impact, given the direct link to lost revenue and reputation. Technically challenging, since computer vision demands precision. Strong on data availability, because images are easy to collect via cameras. Low risk under the AI Act, since the system only assesses product quality and not people.

The success criteria were set in advance: detect at least ninety percent of defects with less than five percent false alarms, and a sorting decision within one second, in real time on the line.

The final solution came after several technical iterations. Xylos worked out different architectures before arriving at the combination that fits this production environment.

A YOLOX model recognizes the position and movement of each bag on the line in real time, so inspection is focused on the right zone. An EfficientAD model picks up the anomalies. That model learns from images of correct seams and flags anything that deviates from them, without every defect type needing to be labeled in advance.

The models run locally on a network of Raspberry Pis, with a central controller. Edge AI instead of a cloud setup, because a production line can't tolerate delay and the cost is a fraction of an industrial PC solution. The system ran in parallel with the existing human inspection for two weeks, so every detection could be compared and fine-tuned. After that, it became the primary check, with the operators as a safety net.

During the test period, the system analyzed thousands of bags, with a detection rate within the target and a false-alarm rate well below the threshold. The adoption test came back unanimously positive from both operators and quality management. Six months into production:

• Defective bags leave the line before reaching the container, sharply cutting product loss.

• Container rejections in export markets have clearly dropped, lowering transport and reshipment costs.

• Seam inspection now needs considerably less human presence, freeing operators to focus on more valuable checks.

• International customers are experiencing fewer quality incidents, strengthening the company's position as a reliable supplier.

Other quality-critical steps are now in line, such as palletizing and wrap finishing. The experience with edge AI made one thing clear internally: advanced AI doesn't have to mean a heavy investment.

Do you only discover production flaws once they reach the customer? Let's find out together where AI can turn that around fastest.