Image Classification & Segmentation

Custom vision pipelines for the visual data that only exists in your industry.

Discuss this See example work

Time to value: Feasibility answer in 2 weeks, production model in 6–10.
Deployment: Can run fully on your infrastructure
Category: Data, Knowledge & Documents

A welder's gloved hands guiding a torch against sparking metal

The manual version of this

Invictt AI layer

Define the classes

Label efficiently

Train and measure

Deploy

A warehouse worker reviewing results on a computer screen at her station

Once it's AI-assisted

01/The problem

General-purpose vision models know dogs and cars. They do not know your weld porosity grades, your fabric defect classes, or the line between acceptable and rejected in your QA manual. That distinction is the entire job.

02/See it in action

What you'd actually see

A demonstration of the mechanism — not a live system reading your data.

Inspection report

Equipment
PPE — compliant
Work zone

Report drafted→ reviewer queue

03/How it works

The operating model

What enters, what the system does with it, and what reaches you.

System

Define the classes

We work from your quality standard to a labelling scheme that is actually separable — this is where most vision projects are won or lost, before a model is trained.

Check

Label efficiently

Active learning puts your experts' time on the ambiguous examples rather than the thousandth obvious one, which usually cuts labelling effort substantially.

Check

Train and measure

Measured on the metric that matters to you — often recall on the rare defect rather than overall accuracy, because the costs are asymmetric.

Integration

Deploy

Exported to ONNX or TensorRT and run wherever it needs to be, in the line or in a batch job overnight.

Read the detail behind each stage

Define the classes

We work from your quality standard to a labelling scheme that is actually separable — this is where most vision projects are won or lost, before a model is trained.

Label efficiently

Active learning puts your experts' time on the ambiguous examples rather than the thousandth obvious one, which usually cuts labelling effort substantially.

Train and measure

Measured on the metric that matters to you — often recall on the rare defect rather than overall accuracy, because the costs are asymmetric.

Deploy

Exported to ONNX or TensorRT and run wherever it needs to be, in the line or in a batch job overnight.

04/What's included

What you actually receive

Concrete deliverables, not a statement of intent.

Labelling scheme design from your quality standard

Active-learning workflow to minimise expert labelling time

Model training measured on your cost-weighted metric

Confusion analysis on the classes that actually matter

ONNX / TensorRT export for in-line or batch deployment

Retraining pipeline for when the product changes

Built with

PyTorch
Segmentation models
Active learning
ONNX
Label Studio

Tooling is chosen per engagement. This is what this kind of build typically uses, not a fixed stack we sell.

05/Example work

All work

What this looks like in practice

Reference implementations, including what went wrong.

Reference Implementation2025

Drafting inspection reports before the inspector gets back to the van

Inspectors were spending their evenings writing up the day's sites from memory and a camera roll. We built an offline-first capture app and a vision model trained on their own defect taxonomy that drafts the report during the visit.

Read it

Reference Implementation2025

A fine-tuned model that replaced an expensive frontier API on the line

A frontier vision API classified fabric defects accurately and cost more than the QA salaries it supplemented. We fine-tuned a small open model on the client's own inspection archive and put it on the line.

Read it

06/Also in this category