Automated visual inspection that finds defects on the line and rejects parts
We install an industrial camera and lighting on your line and train a model on your parts. Inference runs on an edge GPU next to the line. On a defect the system signals the PLC to stop the line, eject or mark the part.
What hurts
An inspector watches thousands of identical parts per shift. Attention drops by the end of the shift and defects slip through.
Defects are found by the customer or in finished goods storage. Then the whole batch needs checking and someone has to trace when the problem started.
There are no defect statistics by type and shift. The process engineer cannot see which operation or setting creates scrap.
What we build
- Line survey and selection of industrial camera, lens and lighting
- Image collection and labelling of your parts, including rare defects
- Defect detection model trained to your tolerances and defect classes
- Edge GPU next to the line so decisions keep up with conveyor speed
- PLC integration over OPC UA or digital I/O: stop, eject, mark
- Results sent to your MES or ERP by batch and shift
- Defect dashboard with a frame of every rejected part, plus Telegram alerts
- Review mode where an operator confirms or rejects borderline cases
What you get
- Installed camera station on the line with tuned lighting
- Trained model, labelled dataset and an accuracy report from your line
- PLC and MES integration with signal documentation
- Defect dashboard and Telegram alerts for the process engineer
- Operator guide and a model retraining plan
How we measure the result
- Share of defects caught on the line rather than by customers
- False rejects on a control batch
- Defects by type, shift and operation
- Time from a defect appearing to engineer response
Questions
What does it cost and what drives the price?
Projects start at $10,000. The price depends on the number of inspection points, line speed, the number of defect types, whether industrial cameras and lighting are needed, and the PLC and MES integration. We give a fixed estimate within 2-3 days after the line survey.
Can we use our CCTV cameras?
For quality inspection, usually not. You need an industrial camera, the right lens and stable lighting, otherwise small defects are invisible. Regular RTSP cameras work for coarse checks, such as whether a label or cap is present.
How many defect images do you need?
It depends on the defect. Rare defects often have few examples, so we collect them during the pilot, use augmentation and methods that learn from good parts. We state the exact data needs after discovery and put them in the plan.
What if the model makes a mistake?
AI sometimes makes mistakes. We set the threshold together with your engineer, trading false rejects against misses. Borderline parts go to manual review and every decision is stored with its frame. Accuracy is measured on your line and written into acceptance criteria.
How long does the project take?
Eight to twelve weeks. Discovery and data collection take the first weeks, then model training, station installation and a shadow run where the system only flags defects. After validation we enable line control and move to SLA support.
Do images leave the factory?
No, inference runs on an edge GPU on the shop floor. Only statistics and frames of rejected parts may go to the dashboard, if you approve it. For restricted sites we deploy everything on your own servers, which also helps with EU AI Act and GDPR documentation.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.