Machine learning
A model that never ships is a research project
The interesting part is rarely the training run. It is the data pipeline before it and getting the result onto a device with a power budget after it.
What we take on
Computer vision, end to end
Datasets
Collection, labelling workflow, and the versioning that lets you answer which data a given model was trained on nine months later.
Training and tuning
Detection and classification models, hyperparameter sweeps, and evaluation against a held-out set you agree on before the numbers exist.
Getting it onto the device
Export to ONNX and TensorRT, quantisation, and measuring what accuracy that trade actually cost rather than assuming it was free.
The console around it
Runs, comparisons and deployments are software problems. We build that side too, which is usually what turns a model into something a team can keep improving.
Where this comes from
Our own lab
We run an internal lab for this work: datasets, training runs, sweeps, versioned models and the evaluation history behind each one. It exists because we needed vision models running on embedded hardware we had designed ourselves, on a compute budget that made the easy answers unavailable.
That is the experience we bring. It is also why the honest limit below is stated rather than left for you to discover.
What we do not do
The limits, up front
We do not sell a model API, and there is nothing here to subscribe to. Every piece of this is per project, built for one client and handed over.
Our shipped work is computer vision. If your problem is language modelling or recommendation, say so in the inquiry and we will tell you plainly whether we are the right people, which some of the time we will not be.
Describe the problem, not the model
What you need recognised, what it runs on, and what happens when it gets one wrong. That is enough for a first conversation.