AI in production
Machine learning, computer vision and anomaly detection — not as an experiment, but integrated directly into your manufacturing processes. Measurable, explainable, maintainable.
Use cases
Predictive maintenance
Vibration, temperature and pressure sensors detect machine anomalies before breakdowns happen. 48 hours' early warning instead of emergency repairs.
Computer vision
Camera-based quality inspection with 95–99% accuracy. Automatic rejection without stopping production.
Optimised shift planning
An ML model combines qualification profiles, machine capacities and historical shift data into optimised shift plans.
OCR & document processing
Automatic classification of delivery notes, test reports and production orders — processing time halved.
Our approach to AI
Data capture & preparation
No ML without good data. We capture the right signals, clean them and structure them for training.
Model development & training
We choose the method that fits your use case — not the trendiest one. A simple anomaly model is often better than a complex CNN.
Integration into existing systems
The model runs where it's needed: on the edge device, in the MES or in the cloud. Alerts via Teams, email or dashboard.
Monitoring & continuous training
Models drift over time. We implement monitoring that tracks model quality and triggers retraining.
Technologies we use
Reference projects
Frequently asked questions
How much data do we need for an ML model?
That depends heavily on the use case. For anomaly detection, 4–8 weeks of normal operating data is often enough. For computer vision we typically need 5,000–15,000 images.
What happens if the model gets it wrong?
We always build in a human in the loop. The model suggests, a person decides — especially in the early phase. Over time, confidence grows.
Can we run AI on-premise, without being forced into the cloud?
Yes. We use edge GPUs or local inference servers. Sensitive production data never leaves the plant.
How long does it take until a first model is in production?
An anomaly detection model can go live in 4–6 weeks. Computer vision typically takes 8–12 weeks because of data collection and training.