Skip to content

AI engineering

Computer vision

Systems that look at what comes off a line or through a door and act on what they see — inspection, counting, sorting, and verification.

What this is

The work, described plainly

Sialkot runs on manufacturing, and manufacturing runs on inspection. Surface defects, dimensional checks, count verification, and packing confirmation are done by eye today, at a rate limited by how many trained people are in the room and how long they have been there.

A camera and a model do not get tired, but they do need conditions. Lighting, mounting, and consistent framing decide whether this works at all — more than the choice of architecture does. We look at the physical setup before we look at models.

Deployment usually runs on-site rather than in the cloud, because sending video off-premise is slow, expensive, and often not permitted.

What you get

What is actually handed over

Concrete deliverables, not phases on a timeline.

Feasibility assessment

Sample images from your actual line, with an honest answer about whether the defect is visible to a camera at all.

Labelled dataset

Built with your team, since only your inspectors know what counts as a reject.

On-site inference

Running on local hardware, so a network outage does not stop the line.

Operator interface

What was flagged, why, and the ability to overrule it — with the overrule recorded as training signal.

Is this you

Signals this is the right fit

  • Inspection is a bottleneck, or quality varies by shift.
  • You need a count or a measurement recorded, not just judged.
  • Defects are visible to a human eye in a photograph.
  • Video cannot leave the premises.

Typical stack

Chosen per project against your constraints, not applied by default. This is what the work usually involves.

  • PyTorch
  • YOLO and Ultralytics
  • OpenCV
  • ONNX Runtime
  • NVIDIA Jetson or equivalent edge hardware

FAQ

Questions people actually ask

  • What if defects are rare?

    Rare defects are the hard case, because there is little to learn from. Anomaly detection — training on what normal looks like and flagging departures — often works better than trying to classify defects you have twelve examples of.

  • Do we need new cameras?

    Often yes, and more often you need better lighting than better cameras. We assess the physical setup first, because no model recovers information the image never captured.

  • Will it replace our inspectors?

    It changes what they do. The system handles volume and consistency; people handle the ambiguous cases and the overrules. Systems designed to remove the human entirely tend to fail on the cases that most needed one.

Start here

Tell us what is slow, manual, or breaking.

Answer a few questions and get a written brief back — scope, proposed architecture, and what it would take to build.