For years, deploying vision-based quality inspection meant an integrator project: a machine-vision specialist came to the plant, chose the algorithms, tuned the parameters and left behind a system nobody on site fully understood. Every product change meant another visit and another invoice. AIVQC was built on the assumption that it does not have to be that way. If your plant has process engineers and an IT department, you already have most of the skills needed to launch AI inspection — what was missing were tools that do not require a machine-learning expert.
This article shows what such a deployment looks like in practice: who you need, what each of them does, in what order the work happens, and at which points it is still worth calling on our support.
Why this is possible at all
In a classic machine-vision project, most of the specialist work goes into building and tuning the model: choosing the architecture, batch size, learning rate and number of epochs, then reacting when results stall or get worse. In AIVQC that part is handled by automatic training in Trainer. You set a time budget and a quality goal, and Trainer fits the parameters to your dataset and hardware, runs successive attempts, improves them based on the results and keeps the best model — not the last one.
The second barrier is infrastructure. AIVQC Server installs like any Docker container — on a Linux server, a Synology NAS or a Windows computer — and the installer guides you through configuration step by step. The third is line integration: the station exchanges a few simple signals with the PLC over Modbus TCP, signals every automation engineer knows from other devices. None of these steps requires knowledge a typical plant engineering department does not have.
Who you need on the team
The deployment does not require a new hire. You need a few people who already work at the plant — each involved at a different stage and to a different extent.
- Process or quality engineer — the most important role. Knows the product and what counts as a defect. Collects and labels images, sets the quality goal, reviews evaluation results and decides whether the model is ready. This person works in Trainer.
- IT administrator — installs Server, gives it an address on the plant network, takes care of backups and updates. Involvement is highest at the start and later comes down to routine maintenance.
- Automation engineer — connects the station to the PLC: trigger signal, reject pulse, heartbeat and healthy signal. Tests the whole logic in simulation mode before the station affects the running line.
- Operators and technicians — run inspection day to day, choose the recipe and react to results. Technicians also collect difficult images that go back to the engineer as material for the next model version.
In a smaller plant the first two roles are often one person. The key point is that none of them needs to be a machine-learning specialist — knowledge of the process, the plant network and PLCs is enough, and you already have it.
Deployment step by step
It is best to start with one station and one product. That is exactly the pilot scope: one Trainer workstation, one inspection station and up to three defect classes for 30 days, with a free extension available. In that time you go through the whole path that later repeats on additional lines.
- Choosing the first case — a repeatable product, a controlled position and a defect that can be shown clearly in an image. A good first case shortens the whole deployment.
- Installing Server — the administrator starts the container on the plant network. Server runs over HTTPS on the local network only and needs no ports opened to the Internet.
- Connecting Trainer and the station — with one-time codes. Each component gets its own identity, and credentials are never displayed or retyped by hand.
- Data and annotation — the engineer imports images or recordings, marks every defect and confirms defect-free images. Trainer makes sure shots of the same part do not end up in both training and test sets.
- Automatic training — time budget, quality goal and start. Training runs in the background, on the engineering workstation or on a GPU computer on the local network.
- Evaluation — testing on images outside training and on video from the target camera, through the same inference path the line uses. Trainer proposes a confidence threshold based on the results.
- PLC integration — the automation engineer tests the signals in simulation mode, then on the stopped line.
- Recipe deployment — the verified package goes through Server to the selected station. The station checks the checksums and keeps the previous recipe for rollback.
- Improvement — NOK and low-confidence images go back through Server to Trainer, and the next model version is created by fine-tuning the previous one.
Where you really save
The biggest difference is no integrator hours — both for the first deployment and for every later change. You pay for a software licence, not for an external specialist's time. How to work out whether it pays off in your case is covered in the article on the return on automating quality inspection.
- No integrator costs for the first deployment — installation, data, training and integration are done by your team.
- No dependency for changes — a new product variant or a new defect type is work for your engineer in Trainer, not an order and a lead time with a supplier.
- Using hardware you already have — if the station camera and computer pass qualification, you do not buy a new set.
- Knowledge stays on site — your people understand how the system works, why it makes the decisions it does and how to improve it.
- Scaling at your own pace — once the first station works, additional lines repeat the same path with the same people.
Honestly: what still takes work
An in-house deployment does not mean the system does everything by itself. Automatic training replaces a machine-learning expert, but it does not replace process knowledge: someone has to collect representative images, label defects correctly and decide which kind of error is worse for your process — an escaped defect or a rejected good part. That decision directly affects the decision threshold.
The image deserves the most attention. A model cannot detect a defect the camera does not show clearly, so imaging conditions are worth checking before you start collecting data — more on that in the article on lighting in visual inspection. You also need to plan time: an engineer with a few hours a week for Trainer during the pilot will go through the whole path far more confidently than someone fitting it in between other tasks.
When our support is worth it
An in-house deployment and working with us are not mutually exclusive. You can start on your own and ask for help in one area only — or run the pilot together with us and deploy additional lines with your own team.
- You are not sure whether the current camera, optics and lighting will show the defect clearly enough.
- The plant lacks an automation engineer to integrate the station with the line PLC.
- You need a formal report with a GO, CONDITIONAL GO or NO-GO recommendation before management approves the investment.
- The defect is unusual and you want advice on how to label it before investing time in annotation.
How to start
The pilot with your own team will be available from the beta, planned for November 2026. You can already sign your plant up through the contact form — choose the topic "Beta — pilot with your own team" and describe the product, the defect and how you inspect it today. We quote the licence individually, and if it turns out along the way that you prefer a deployment with our support, you can switch paths at any time.