Computer vision
The model works on the sample images. The question is whether it works in the car park, at six in the evening, in the rain.
Conditions are the whole problem
Vision models fail in ways that are almost never about the model.
They fail because the camera moved. Because the lighting changed with the season. Because someone put a sticker on the thing being photographed. Because the training data was collected on one device and production runs on another, three years older, with a dirtier lens.
Getting a vision model to work on curated data is a solved problem you can do in a weekend. Getting it to work reliably on whatever your environment actually produces is the engagement.
We spend most of our time on the second thing.
What we build
- Visual inspection
- Defect detection, quality grading, and compliance checks on production lines or in the field.
- Object detection and counting
- Locating and tallying things in images or video, where "things" is domain-specific and generic models won't cut it.
- Document and form understanding
- Layout-aware extraction from scans, photos, and PDFs. Where the information is as much in the position as the text.
- Classification and sorting
- Categorising images at volume, against categories that matter to your business rather than to ImageNet.
How we approach it
- We audit the capture conditions before the data
- What camera, what angle, what lighting, what varies. Half the accuracy problems in vision projects are fixable with a better mounting bracket, and that's a conversation worth having before you pay for a model.
- We build the hard test set first
- Not the clean images — the blurry ones, the odd angles, the edge cases someone will encounter in week two. If it works on those, the average takes care of itself.
- We're realistic about labelling
- Vision projects need labelled data and labelling costs real time. We'll tell you how much you need, help design the labelling process, and look for ways to reduce the requirement — synthetic data, augmentation, transfer learning — before we ask you to fund a labelling exercise.
- We plan for the deployment target
- A model that needs a GPU is a different proposal from one that runs on a phone or an edge device. That constraint shapes the architecture, so we settle it at the start.
Before you ask.
Often, yes, with the right architecture. It's a constraint we design around from the beginning rather than discovering at the end.
Machine Learning Models
A model that scores 94% in a notebook and a model that earns its keep are different achievements. The distance between them is where this work lives.
Intelligent SystemsDocument Processing
Somewhere in your business, a person is reading a PDF and typing what it says into a form. Possibly right now.
Data & CloudData Engineering
Nobody has ever asked us for data engineering. They ask for the dashboard, the model, or the AI feature — and then we find out why it isn't working.
Tell us what you're trying to build.
A 30-minute call, no charge and no pitch deck. Describe the problem and we'll tell you how we'd approach it, roughly what it costs, and whether we're the right team for it. If we're not, we'll say so.
30 minutes · No charge · No deck