Computer vision that works on your images
Gjethe builds computer vision systems: object detection and counting, visual quality inspection, OCR and document data extraction, image classification and tagging, video analytics, and visual search. We evaluate modern vision and multimodal AI models against your own images, train custom models where needed, and deploy them in the cloud, on edge devices or inside your apps, with privacy built in.
- Free first consultation
- Written quote for your project
- Flexible payment milestones
At a glance
- Tasks
- Detection, inspection, OCR, classification, video
- Models
- YOLO, vision transformers, multimodal LLMs
- Deployment
- Cloud APIs, edge devices, mobile apps
- Privacy
- Anonymisation and on-device processing options
Definition
What is computer vision?
Computer vision is the field of AI that lets software interpret images and video: recognising objects, reading text, measuring, classifying and detecting changes or defects.
Cameras are cheap and everywhere, but most image data is never used, because looking through it takes people too long. Computer vision can count, check, read and sort images at scale. It works best on well-defined tasks with consistent conditions, and its accuracy must be measured on your real images, not on impressive demos.
Who is it for?
Manufacturers
Automated visual inspection for defects, assembly and labelling.
Document-heavy operations
Data extracted from invoices, forms, IDs and delivery notes.
Retail and e-commerce
Product tagging, visual search and shelf monitoring.
Construction and agriculture
Progress tracking, safety checks and crop monitoring from photos and drones.
What's included, and what you keep
What's included
- Use-case definition and accuracy targets
- Image collection and labelling plan
- Evaluation of ready-made and multimodal models
- Custom model training where needed
- Object detection and counting
- Defect and anomaly detection
- OCR and structured document extraction
- Image classification and tagging
- Video stream analysis
- Deployment to cloud, edge or mobile
- Monitoring and review workflows
What you receive
Handed over at the end. Yours to keep.
- Evaluation report on your images
- Trained models and training code
- Deployed inference service or edge application
- Labelled dataset and labelling guidelines
- A review interface for uncertain results
Ready-made model, multimodal AI or custom training?
Many tasks can now be handled by general multimodal models that understand images and text, with no training at all, such as reading a delivery note or describing a product photo. They are quick to try but can be slower and more expensive per image, and less precise on specialised tasks.
For high-volume or specialised tasks, such as detecting a specific defect on a production line, a smaller custom model trained on your images is usually faster, cheaper to run and more accurate. We test the options on your data before recommending one.
Typical computer vision tasks
| Task | Example |
|---|---|
| Detection and counting | Counting items on a pallet or people entering a space |
| Inspection | Finding scratches, missing parts or wrong labels |
| OCR and extraction | Reading invoice fields, serial numbers or meter readings |
| Classification | Sorting photos by product type or condition |
| Segmentation | Measuring areas such as damage or crop coverage |
| Video analytics | Detecting events in camera streams |
Your images decide accuracy
Lighting, camera angle, resolution, motion and background all affect results. A model that works well on clean photos may struggle on a dusty production line at night. We collect representative images early, including the difficult cases, and measure accuracy on them. Where conditions can be controlled, a better camera position or lighting often improves results more than a better model.
People stay in the loop
No vision system is perfect. We design workflows where uncertain results go to a person for review, and those reviews become new training data. This keeps errors visible and lets accuracy improve over time.
Privacy and cameras
Images often contain people, faces, licence plates or personal documents. We design for data minimisation: processing on the device where possible, blurring or discarding personal details, limiting storage and restricting access. We don't build facial recognition to identify individuals, and your legal advisers confirm the rules that apply to cameras in your workplace or premises.
What helps us start
- Example images or video, including difficult cases
- What a correct result looks like
- The accuracy you need, and the cost of mistakes
- Where the images come from and how many
- Where results must be sent
- Constraints on cloud processing or storage
From pilot to production
A vision pilot often works on a few hundred images in good conditions. Production means thousands of images a day, new product variants, camera faults and seasonal light changes. We plan for this from the start: monitoring of accuracy and confidence, alerts when inputs look unusual, and a simple way to add new examples when something changes.
How computer vision works, step by step
Every step has a clear output you can review. Timing is agreed in your quote.
- 1
Define
Task, accuracy target and the cost of each type of error.
- 2
Collect
Representative images gathered and labelled.
- 3
Evaluate
Ready-made, multimodal and custom options tested on your data.
- 4
Build
Model, pipeline and review workflow developed.
- 5
Deploy
Cloud, edge or mobile deployment with monitoring.
- 6
Improve
Reviewed cases feed back into training.
Tools, and what shapes your quote
Pricing is based on your project and agreed in your quote. These are the things that move it.
| Factor | Why it matters |
|---|---|
| Task complexity | Detecting subtle defects is harder than counting obvious objects. |
| Data and labelling | Collecting and labelling images is often the largest part of the work. |
| Accuracy targets | Higher accuracy needs more data and evaluation. |
| Deployment | Edge devices and real-time video need optimisation. |
| Integration | Connecting results to your systems adds development. |
| Privacy requirements | Anonymisation and on-device processing add work. |
| Area | Tools |
|---|---|
| Models | YOLO, vision transformers, SAM for segmentation, multimodal LLM APIs |
| OCR | Tesseract, PaddleOCR, cloud document AI services |
| Training and labelling | PyTorch, Ultralytics, Label Studio, Roboflow |
| Deployment | ONNX, TensorRT, NVIDIA Jetson, Core ML, TensorFlow Lite |
Common mistakes to avoid
What we see go wrong most often, and how a good project avoids it.
Testing on perfect images
Real conditions are messier. Test on the images you'll actually get.
Ignoring the camera set-up
Better lighting and positioning often beat a better model.
No human review
Uncertain results need a person to check them.
Collecting more personal data than needed
Blur, crop or discard what the task doesn't need.
Frequently asked questions
Can you extract data from invoices and forms?
Yes. We combine OCR or multimodal models with validation rules, and send unclear cases for review.
Do you build facial recognition?
We don't build systems that identify individuals by their face. We can detect and blur faces to protect privacy.
Can it run on a device without internet?
Yes. Models can be optimised for edge devices such as NVIDIA Jetson or phones.
How many images do we need?
It depends on the task. For ready-made or multimodal models, a small test set may be enough; custom models usually need hundreds to thousands of labelled examples.
Can you inspect products on our production line?
Yes, with camera, lighting and model designed together, and tested under real line conditions.
Can you analyse video from our existing cameras?
Often, if resolution, angle and lighting are suitable. We assess sample footage first.
Who labels the images?
We can label them with guidelines you approve, or train your team to label consistently.
Who owns the trained model?
You do, along with the labelled data and code, once the project is paid.
How is computer vision priced?
It is quoted per phase, starting with an evaluation on your images, then development and deployment.
Can you count products or stock from photos?
Yes. Counting and detection are common, reliable tasks when images are reasonably consistent.
Can it read handwriting?
Modern models read many types of handwriting reasonably well. Accuracy depends on the writing, so we test on your samples first.
How fast can images be processed?
It depends on the model and hardware. Edge devices can process camera frames in real time for many tasks; cloud processing suits batches of photos and documents.
Can you combine vision with our existing software?
Yes. Results can be sent to your ERP, warehouse system, CRM or a dashboard through an API.
Related services
Computer Vision by industry
Put your images to work.
Tell us what you need. You get an honest assessment, a clear plan and a written quote. The first consultation is always free.
- What happens next
- 1You tell us what you need
- 2We aim to reply within 24 hours with questions
- 3You get a written quote for your project