Example system · Manufacturing

Visual quality inspection

Inspect every part, at line speed.

Example system Manufacturing

Inspect every part, at line speed.

  1. 01Sense
  2. 02Learn
  3. 03Predict
  4. 04Act

Sense. A camera and controlled lighting at one station, triggered by the line itself, so every part is photographed the same way.

Illustrative system. Simulated data, not client results.

Inside the engineering

The problem

People inspecting parts catch most defects early in a shift and fewer by the end of it. Sampling misses whatever falls between samples, and by the time a pattern is noticed, a whole batch has been made.

Why it is hard

  1. 01Defects are rare, so there are few examples to learn from.
  2. 02Lighting, dust and new batches of material change how parts look.
  3. 03The decision has to be made in the time it takes a part to reach the gate.
  4. 04Quality teams need to see why a part was rejected, not just that it was.

What we build

From signal to decision.

  1. 01

    Sense

    Industrial cameras, lighting and a trigger from the line encoder, mounted so the image never changes between shifts.

  2. 02

    Learn

    A vision transformer fine-tuned on your parts, with an anomaly detector beside it for defect types it has never seen.

  3. 03

    Predict

    A verdict, a heat map of where the defect is and a calibrated confidence for every part, on an edge computer next to the line.

  4. 04

    Act

    Reject gate control, a review queue for uncertain parts, and defect trends by cavity, tool and shift, sent to the people who own them.

Before we start

What needs to be in place.

  • A station on the line with room for a camera and stable lighting.
  • A few hundred labelled images per defect type, which we help you collect.
  • A quality engineer to agree what counts as a reject.

How the project runs

From survey to live.

Fixed price once the specification is agreed. Every repository, model and account is yours from the first day.

  1. Survey1–2 weeksLine speed, defect types, and where a camera can go.
  2. Capture2–3 weeksInstall the station. Collect and label images with your quality team.
  3. Train3–5 weeksModel and thresholds, run in shadow mode beside human inspection.
  4. Go live1–2 weeksGate control, the review queue and trend reports, handed over.