Computer vision that checks, reads and counts from camera images
Every item checked by a camera, with a photo kept for each decision.
Use case · A packaging manufacturer with printed cartons
The inspector that never blinks
A composite of the projects we take on: the problem, what we built, and what changed. Pick an act to jump the film to it.
Who it is for
Is this you?
- Factories that inspect products by eye
- Teams typing data from forms, invoices and IDs
- Sites that count people, vehicles or stock from video
What you get
- Feasibility test on your own images
- Labelled dataset and trained model
- Model running on edge devices, phones or the cloud
- Integration with your line, app or ERP
- Review screen for cases the model is unsure about
- Monitoring and a retraining plan
Built with
- Python
- PyTorch
- OpenCV
- YOLO
- ONNX
- TensorFlow Lite
- NVIDIA Jetson
- Label Studio
What we cover
Computer vision, in detail
Pick one part or the whole set. We scope only what your problem needs.
Inspection
- Defect and damage detection
- Print and label checks
- Assembly and fill-level checks
- Measurement from images
Documents
- OCR for printed and handwritten text
- Invoice, receipt and form extraction
- ID and document scanning in apps
- Table and stamp detection
Detection and tracking
- Object detection and counting
- Real-time video analysis
- Shelf and stock monitoring
- Safety gear and zone alerts
Privacy and deployment
- On-device processing
- Automatic face and plate blurring
- Edge, mobile or cloud inference
- Data handling agreed before we start
How it runs
From first call to computer vision in production
You see working software early and can change direction while it is still cheap.
- 01
Test on your images
We start with a short feasibility test on real images from your line, app or cameras. - 02
Collect and label
We set up capture, label the images with your experts and agree what counts as a defect. - 03
Train and fit
We train the model, run it on the target hardware and tune it for your speed and lighting. - 04
Run beside people
The model runs next to your inspectors first. We widen its role as the results hold up.
Questions
Computer vision: what people ask
How many images do you need?
It depends on how varied the items and defects are. The feasibility test tells us, and we often start with a few hundred labelled images per defect type.
Do we need special hardware?
Not always. Many jobs run on an industrial camera and a small edge computer, and some run on a phone. We recommend hardware after the feasibility test.
What happens when the model is unsure?
It sends the image to a review screen where a person decides. Those decisions become new training data, so the model improves over time.
Planning computer vision work?
Tell us where things get stuck today. We will reply with a plan, not a pitch.