Questions and answers

It works best in mid-sized manufacturing plants starting with one to a few stations. In offline mode, Server runs on the Trainer computer and serves up to three stations, so you do not need a separate server. Larger plants install Server as a Docker container and add more lines.

Ones that can be shown clearly in an image: missing components, incorrect assembly, flash and deformation, cracks and damage, label and print errors, and contamination. We describe each of these in Applications. If a defect is not clearly visible in the image, imaging has to be improved first — the camera, optics or lighting.

Most often plastics processing, automotive (Tier 2 and 3 suppliers), metal components and packaging — see Industries for details. Whether AIVQC fits, however, depends on the nature of the defect and the imaging conditions, not the sector itself.

AIVQC is software that runs on the plant's own computers and cameras, after they pass qualification. Training happens on site, approved model versions are published to selected stations with the option to roll back, and permission to change thresholds depends on the role. We compare four approaches and list the questions worth asking vendors in Smart camera or AI system?.

The pilot is the full platform on one station: one Trainer workstation, one inspection station and up to three defect classes for 30 days, with a free extension available. You train a model on images of your own product and see the result on your line before you decide on deployment. A pilot with our support ends with a report recommending GO, CONDITIONAL GO or NO-GO. Step by step: AI quality-inspection pilot.

Yes — from the beta, that is from February 2027. You install Server, connect Trainer and the station, and automatic training prepares the model. If you want to start earlier or prefer us alongside you, we will run the pilot together with you. You can sign up for the beta through the contact form.

We do not use one threshold for every project. Acceptance criteria are agreed before the pilot based on the risk and economics of the specific process. How to weigh false accepts against false rejects is covered in the article on decision thresholds.

You choose how to deploy. We can prepare an offer for a full production deployment — licence, hardware qualification, configuration and integration — or provide the software alone if your own team deploys it.

Yes. If you have your own process engineers and IT staff, you have everything you need: Server installs like any Docker container — or, at a small site, as a service on the Trainer computer — and automatic training needs no machine-vision expert. You pay for a software licence, not for integrator hours — usually the biggest saving. Contact us for a quote.

A quality or process engineer who knows the product and its defects, IT staff to install Server, and an automation engineer to connect the station to the PLC. None of them needs to be a machine-learning specialist. Details are in the guide Deploying with your own team.

We can help in one area — for example choosing the camera and lighting, or connecting to the PLC — or run the pilot together with you. Deploying with your own team does not mean you are left on your own.

Yes, if they pass qualification with the target camera, model, recipe, I/O and required cycle time. A processor or camera name alone does not guarantee the result. How to assess it is covered in Retrofit or new station?.

There is no single number — it depends on how much the product and the defect vary. More important than the count are representative images and an independent test set the model never learns from. Details in How many images do you need to train an AI inspection?.

The model is evaluated on an independent test set: we measure detection quality, false accepts and false rejects, and decision time on the target hardware. Acceptance criteria are set before the pilot, and an approved version reaches the line only once it meets them.

Images from the line — for example NOK or low-confidence ones — can return to Trainer through Server. There you label them, retrain the model and publish a new recipe version to selected stations. If the new version does not perform, you can roll back to the previous one. Your own engineer does this, without an order to an integrator.

Four: OK, NOK, REVIEW and ERROR. NOK means a defect was found — with a PLC connected, the station sends a reject pulse. REVIEW means there is no reliable decision, for example when the product is partly out of frame; a person checks that part instead of the system guessing OK. ERROR means the station could not run the inspection (camera, model or timing) — and it is never treated as OK. The station interface is currently in English.

Over Modbus TCP, with a few signals: trigger (the product has reached the camera), line running, a reject pulse for a NOK product, heartbeat and station ready. Simulation mode lets you check the whole logic together with your automation engineer before connecting the real line. AIVQC is not a safety device — the PLC reacts to a missing heartbeat or ready signal and decides whether to stop or divert the line. Details on the AIVQC Production page.

It changes their work. The station checks every part with the same logic on every shift, and people move from looking at everything to handling exceptions — parts marked REVIEW, defect analysis and process improvement. The operator starts inspection with one button and sees the result with its decision time.

No. Training runs on your own computers, and image analysis runs at the AIVQC Production station. Inspection and package publishing work on the plant network, also without Internet access. More in Why production images do not have to leave the plant.

Server runs as a Docker container on an existing server, NAS or Windows computer, and at a small site as a service on the Trainer computer. It needs no ports open from the Internet — it runs over HTTPS on the local network. Trainer, each station and the administrator have separate identities and roles. Server makes daily encrypted backups and updates in a maintenance window with the option to roll back. The optional connection to AIVQC Relay is outbound only and is used for the licence and signed updates; without it you activate offline. The full description of connections, data and roles is on the Security and IT page.

The AIVQC camera looks at the product, not at people — ideally frame it so that only the part is in the image. Images showing people, badges or documents should not go into the training set and should be deleted. If people regularly appear in the frame, include the camera in the plant's video-monitoring rules (employment law, GDPR). Images stay in your infrastructure, which makes them easier to control.

Inspection does not depend on our cloud. The Production station uses the last locally stored, verified version and keeps running even without a connection to Server — only downloading new versions is lost. Images, datasets, models and inspection history are on your computers, not ours, and activation can be done offline. Details on the Security and IT page.

Your line keeps running. We commit that in that case AIVQC will keep receiving security updates for at least 2 years and licences will be made perpetual. Images, datasets, models and inspection history are on your computers, and models are stored in the open ONNX format.

Product quality inspection is, as a rule, not a safety component, and it is a safety function in a product covered by EU harmonisation legislation that determines high risk. AIVQC is not a safety device — the PLC program decides whether to stop the line. If you inspect parts that are critical to people's safety, assess the classification individually with a lawyer. This answer is not legal advice.

It provides the data an auditor asks for: validation results on an independent test set, inspection history with images, and which model, recipe and configuration were responsible for each result. The procedures the standards require — measurement system analysis for OK/NOK judgements or checking error-proofing with a master part at the start of a shift — are organised by the plant, and AIVQC supplies the records and images for them.

We quote individually — a pilot and deployment with us, or the software alone if your own team deploys it. You do not pay for a cloud or for the number of images processed, because training and inspection run on your hardware. How to work out the return using data from your own line is covered in How to calculate the return on automating quality inspection.

Often, but it depends on the programme. Dig.IT grants for manufacturing SMEs in Poland fit best — usually as part of a wider digitalisation project. The robotisation tax relief generally covers a vision system only when it works in a cell with an industrial robot. Details and caveats: Funding a vision inspection system in Poland. If you are planning an application, we will help describe the solution.

No. Automatic training chooses the model and parameters, improves them between attempts and picks the best result, while the Overview screen suggests the next step. What matters is knowing the process and the defects — you decide what counts as a defect. A manual mode with the full parameter set is available for advanced users.

During the pilot we jointly define what images are needed and help plan how to collect them. In ongoing operation, production data is collected by your team under the agreed imaging conditions.

No. During the pilot you train as often as you need — every annotation fix or new batch of line images can go straight into the next automatic training run.

Yes. Annotation, evaluation and deployment do not need a GPU. Training can run on the CPU (slower, for experiments) or on a GPU training worker on the local network.

No. Projects, datasets and training stay on your computers, and exchange with the lines goes through the local AIVQC Server on the plant network.

Yes — an ONNX model with its class file can be imported, evaluated and deployed if it meets the AIVQC Production detection contract. It goes through the same acceptance criteria as a model trained in Trainer.

No. A typical deployment is a Docker container on an existing Linux server, a Synology NAS or a Windows computer with Docker Desktop. At a small site — up to three stations — Server can run in offline mode as a service on the Trainer computer.

Not for daily work. Package publishing and inspection run on the plant network. The outbound connection to AIVQC Relay is for licensing and updates, and without it you can activate offline with a file.

No. It stores only what is operationally needed: packages on their way to stations, acknowledgements, history and reports. Line images wait on Server only until they are imported in Trainer.

Production uses the last locally stored, verified version and continues inspecting. The connection to Server is needed to publish new versions, not for ongoing operation.

Production keeps a local copy of the last verified package and continues inspecting offline. Only downloading new versions is affected, until the connection returns.

Yes — recipes let a single station handle several product variants, provided the station has been qualified for each of them.

Not directly. The station sends a reject pulse, a heartbeat and a healthy signal, and the PLC program decides to stop or divert the line when the station reports a fault.

Only if the expert allows it. The threshold from the recipe approved in Trainer is the starting value. The station expert decides whether operators may change it, a technician may adjust it within an allowed range around that value, and the expert can change it freely. A permanent change for every station is a new recipe version from Trainer.

A camera connected to the station computer, or a network camera over an RTSP or HTTP stream. Camera and lighting choices are made during the pilot, because they decide image quality.

That is exactly what the pilot assesses — qualification includes checking whether the camera angle and position show the assembly point clearly enough. If not, the report states that as a limitation.

It depends on how similar the variants are. AIVQC Production uses recipes, so different variants can have a separately configured inspection on the same station.

Not necessarily — it depends on how visually different the variants are. This is decided during data preparation in Trainer, before the actual training run.

Retrofit-first means we first check whether the existing camera and its position allow that. If not, the pilot report states exactly where the limitation lies.

That is exactly what validation measures — the acceptance threshold is agreed with you before the pilot, based on the risk and economics of the process, not a single universal threshold.

It's worth flagging during the initial consultation — it helps us plan the dataset to cover that scenario instead of missing it.

That depends on the contrast between them under the available imaging conditions — exactly what the camera, optics and lighting qualification checks at the start of the pilot.

Yes, if they are visible to the camera and optics used, at the resolution they offer. If they require specialised imaging (e.g. X-ray), that falls outside standard visual inspection.

The base scenario is verifying presence, orientation and general legibility. Reading specific content (e.g. a batch number) is an additional scope we agree on individually during consultation.

That is a typical case where lighting quality decides the result — qualification at the start of the pilot checks whether the current imaging setup handles it.

That depends on the contrast between them in the available images — exactly what the pilot assesses before a GO or NO-GO recommendation.

That depends on camera resolution and field of view relative to product size. If the current hardware cannot support it, the pilot report will flag it as a limitation to consider.

It depends on how much colour affects defect visibility. Recipes in AIVQC Production let you configure inspection separately for different variants on the same station.

Not always — it depends on how much the material's appearance changes. This is one of the questions assessed during data preparation in Trainer.

Server's publication history and Production's local inspection history record which model and configuration produced every result. The exact scope of compliance with a specific OEM's requirements is agreed individually.

Yes, provided the station has been qualified for each variant — recipes let you assign a separate inspection configuration to each one.

It does if lighting hasn't been matched to it — that's one of the first things we check during camera, optics and lighting qualification before the pilot.

It depends on whether the porosity is visually apparent at the available resolution. Defects requiring X-ray imaging fall outside standard visual inspection.

That is exactly what the pilot measures — decision time on the target hardware is one of the qualification report's criteria, alongside detection quality.

The base scenario is verifying presence, orientation and general legibility. Reading specific content is an additional scope agreed individually.