Context

Vision AI is the application for quality management — on the same foundation as Process AI, which carries maintenance planning and maintenance. You operate it in the platform; how it works technically is under the architecture.

The Problem

Manual visual inspection covers a sample, not the production. It depends on how alert someone is that day, and it often finds defects late — at the next process step, in final assembly, or at the customer. Every one of those steps makes the same defect more expensive.

Then there is what never gets checked at all: tool wear, material flow and slow deviations would need continuous observation, and nobody on a shift has time for it.

Who It Is For

Manufacturing with a recurring visual inspection and at least one camera on the line. The more clearly you can name which defects should be found, the faster the pilot pays off.

What We Do

We build a small, self-contained model for your parts — pure software on the existing cameras, no new hardware. It works on the same principles already in use on real machine signals: capture features, compare across cycles, then classify. Where confidence is insufficient, the system flags the case for a human rather than guessing.

What You Get

Visual inspectionAutomated inspection of your parts on the cameras you already have
Defect detectionDefects and dimensional deviations are found and marked in the image
PASS / FAILOne clear decision per inspection, not a probability to interpret yourself
Inspection historyEvery inspection stays traceable — stored locally, in your installation
Operator feedbackYour people can confirm or correct a result
Model versioningModels are versioned rather than overwritten — a given state stays reproducible
Comparison reportHit rate and false positives against your current inspection, measured on your parts

What Is Demonstrable Today

In production useExisting — Vision AI runs today in a real manufacturing environment, not as a demonstration
No new hardwareExisting — software on existing cameras, no intervention in the equipment
Nothing leaves the plantExisting — inspection, history and model on your hardware, air-gap capable
Flag rather than guessExisting — where confidence is insufficient, a human decides

How the Pilot Runs

  • Week 1 — Scoping: Which defects should be detected, which cameras are already there, what does a good part look like?
  • Weeks 2–4 — Build: Connecting to the existing images, a first model on your parts, integration into the platform.
  • Week 5 — Parallel run and handover: The model runs alongside your current inspection. You receive the trained model, the comparison report, and a recommendation on whether production use is worth it.

Effort

DurationAround five weeks, depending on the number of defect types and cameras
PriceA fixed price, set after scoping — before the project starts, in writing
PrerequisiteAt least one existing camera at the station to be inspected

What We Do Not Promise

No accuracy figure before we have seen your parts. No numbers extrapolated from other plants. What you get instead is the comparison report from your own line — and the recommendation to stop, if it does not pay off.

The Next Step

A thirty-minute conversation about the inspection task: what should be detected, what is on the line, how it is checked today. After that you know whether a pilot makes sense — and we know what it costs.

Discuss Vision AI →

Frequently Asked Questions

Do we need new cameras or hardware?

No. It is pure software running on cameras already installed on your line. Where a camera is missing, we name that during scoping — before the quote, not in the final invoice.

How is accuracy validated?

The model runs alongside your existing inspection during the pilot. You receive a comparison report with hit rate and false positives, measured on your own parts, as the basis for deciding on production use. We do not quote accuracy up front — and distrust anyone who does.

What happens with uncertain cases?

Where confidence is insufficient, the system flags the case for a human rather than guessing. That is a design principle, not a shortcoming: better an open question than a wrong rejection.

Do our images leave the site?

No. Inspection, history and the model run entirely on your hardware. No cloud connection is required — the system works air-gapped.

Who owns the trained model?

The model trained for your parts, your configurations and your image data are yours. The underlying technology — Vision AI, the platform and Henri Neuronal Core — stays with SynapSync; you receive usage rights for the agreed perimeter.

Related