Now booking Q1 2027
Now booking Q1 2027Start a project

22:00 — this craft's hour on the studio wall

Computer vision development for startups and product teams

We build computer vision systems that read images and video from your cameras, scanners or phones: spotting defects, counting and tracking objects, reading text and labels. Models are trained on your footage, tested in your conditions and sized for the hardware they will run on.

Start a computer-vision projectEstimate the cost

What we help with

  • Visual inspection done by eye

    People checking parts, packaging or produce for defects get tired, and tired checks drift. A vision model can flag likely defects for review and keep a record of every check.

  • A demo model that falls apart on site

    Models trained on public datasets often miss in real lighting, angles and backgrounds. We collect and label examples from your own cameras, awkward cases included.

  • Video nobody has time to watch

    Counting, tracking and alerts on footage from shops, sites or roads, so people only look at the moments that matter.

  • Paper and screens that need reading

    OCR for documents, labels, meters and serial numbers, with checks that catch misreads before they reach your system.

What you get

  • A pilot on your own images that measures how well the task can be done
  • A labelled dataset, and the guidelines used to label it
  • Trained models with an evaluation broken down by condition: lighting, angle, camera
  • Deployment to the cloud, a phone or an edge device such as an NVIDIA Jetson
  • A simple review screen for flagged results
  • Model weights, training code and data, in your accounts

What we build with

Proven, open tools, exported so they run where you need them.

YOLO detectors
fast object detection and segmentation that runs in real time on modest hardware.
OpenCV
the camera handling, calibration and image processing around the model.
ONNX and TensorRT
one trained model exported to run on servers, phones or edge devices.
NVIDIA Jetson
compact computers on site, for when footage can't leave the building or the network is unreliable.

How an engagement runs

Most projects move through these in order. You can stop after any of them and keep what was made.

  1. Scope sprint

    2 weeks · fixed fee

    Turn an idea into a plan you can trust before you commit to a build.

    We review sample footage and your cameras, check feasibility and quote the build.

    • A clickable prototype
    • A list of risks and how to retire them
    • A fixed quote for the build
    • The fee is credited against the build

    Sample price: from $3,000

    Start with a scope sprint

  2. Build

    6–12 weeks

    We design and ship the product, with something real to try every Friday.

    Data collection, labelling, training and deployment, with a weekly evaluation on footage the model hasn't seen.

    • Weekly builds you can open on your phone
    • Design, engineering and launch
    • Code, models and data handed over
    • 30 days of fixes after launch

    Sample price: typically $15,000–60,000

    Plan a build

  3. Ongoing team or ML pilot

    monthly

    A small senior team that keeps improving a product you already run.

    A pilot first when feasibility is unclear; after launch, retraining as products, sites or cameras change.

    • A fixed team, no rotating juniors
    • A monthly plan and a weekly demo
    • Pause or stop with a month's notice

    Sample price: from $4,000 a month

    Talk about a team

Prices here are samples, to give a sense of scale. A scope sprint ends with a fixed quote for your project. For a rough range before you write, try the cost estimator.

Common questions

How many images do you need?

It depends on how varied the scene is and how subtle the thing you are looking for is. We start with whatever you have, train a first model and use its mistakes to show which extra examples would help most.

Can it run without sending video to the cloud?

Yes. Models can run on an edge device on site, such as an NVIDIA Jetson, or on a phone, so footage stays where it was captured. We size the model to the hardware during the pilot.

Do you work with our existing cameras?

Usually. We check resolution, frame rate, placement and lighting early, and tell you if changing a camera would improve results more than a better model would.

Who labels the training data?

We set up the labelling tool and guidelines and do the first rounds ourselves, so the rules are right. For large volumes we can train your team or bring in a labelling service you approve.