AI defect detection on the production line

Defects caught on the line, not by the customer

We train the model on footage of your product. What counts as a defect is agreed with your process engineer before launch, not after the first complaint. A defect on the belt — the supervisor gets an event with a frame and a timestamp, and the log builds a picture by shift.

Where defects slip through

  • A defect was found in the warehouse a week later. The batch is already with the customer: a return and an awkward conversation are next.
  • By the end of the shift the inspector is tired and, at line speed, misses what they would have spotted in the morning.
  • Which shift and which area produce more defects — nobody can say: there are no statistics by defect type.
  • There are cameras over the line, but footage is opened only after a complaint.

What changes after launch

  • 01Defects are caught on the line while the batch can still be stopped
  • 02Defect statistics by type, area and shift
  • 03Inspectors review borderline cases, not the whole flow

How the system works

  1. 01

    We agree on what a defect is

    Together with your process engineer we fix the defect types, camera angle and line speed. Borderline examples are settled on photos before labelling, not after launch.

  2. 02

    The model learns your product

    We film good and defective items on your line and label the dataset. The model learns from your footage, not from someone else’s pictures.

  3. 03

    We set thresholds

    We decide what matters more: not missing a defect or not bothering the supervisor with false alarms. The trade-off goes into the pilot contract.

  4. 04

    The supervisor is alerted

    An event with the defect type, frame, time and area goes to Telegram, the event log or your MES.

  5. 05

    Acceptance by metrics

    The share of correct detections and false alarms per shift are checked on a control set. The criteria are agreed before the start.

Which scale is yours?

One task, three different projects. Pick the one that looks like you and press “I want this” — your request arrives tagged, and we send an estimate in 1–2 days.

Small business

common start

One line or inspection point and 1–2 defect types behind most of your returns.

What we build

  • 1 line or inspection point, 2–4 cameras
  • 1–2 defect types to start
  • Labelling and training on footage of your product
  • Supervisor alerts and a log by shift
  • Pilot report with detection metrics
Integrations
Telegram, your IP cameras, event log
Where it runs
Edge unit next to the line: footage never leaves the plant
Timeline
Guideline: 3–6 weeks for the pilot

Mid-size business

common start

Several lines or inspection points, a dozen defect types and a process engineer who needs statistics, not impressions.

What we build

  • Several lines and inspection points
  • New defect types added as footage accumulates
  • Events passed to MES or ERP (1C) and linked to batches
  • Defect dashboard by type, area and shift
  • Support contract: retraining when the product, packaging or lighting changes
Integrations
MES, ERP (1C), Telegram, line cameras
Where it runs
Edge units at the lines and a server on site
Timeline
Guideline: 2+ months, starting with one inspection point

Enterprise

Several plants, shared quality standards and your own video surveillance system that every connection has to go through.

What we build

  • One defect catalogue for all sites
  • Integration with your VMS instead of a second camera network, where the angle works
  • Roles and permissions for QC, process engineers and management
  • Event and action log for audits
  • Service under an SLA
Integrations
VMS, MES, ERP (1C), quality management systems
Where it runs
Your servers or a closed perimeter: footage never leaves it
Timeline
Guideline: 3+ months, starting with a pilot on one line

Pilot from RUB 500,000. Beyond that we price by task and hours.

Questions

Do you guarantee that no defect gets through?

No — and do not trust anyone who does. Target detection metrics and an acceptable false alarm rate go into the pilot contract and are checked on a control set.

Do we need special cameras?

Existing cameras are often enough with decent lighting and angle. Small defects on a fast belt may need a different camera or extra lighting — that becomes clear during the assessment. Hardware is a separate line in the estimate.

How is this different from regular CCTV?

CCTV records a picture. This produces events: defect type, time, area. They feed shift reports, so nobody has to scroll through footage by hand.

What happens when we change the product or packaging?

We retrain the model on new footage. If your range changes often, we sign a support contract and plan retraining in advance.

A quote in 1–2 days

Name the process that eats time. You get a pilot range, not a 40-slide deck.

Or message us