The framework most current model work is published and trained in, and the one with the shortest path from paper to running code.
What we reach for first
Custom models, fine-tuning, and computer vision where an off-the-shelf API does not fit the domain. That normally means the data is specialised and belongs to the client — a defect pattern, a document layout, a signal nobody else holds a corpus of. Where that is true, a small model trained on it beats a large general one and costs less to serve.
Problems a hosted model already solves. Training something to do what an API does is a research budget spent on a solved problem. We would rather lose the more interesting project than bill a research budget for an outcome an API already delivers, and we will put the comparison in front of you before either of us commits to it.
A held-out evaluation set defined before training, and the metric agreed with whoever will act on the output. Runs are reproducible from a recorded data version and a seed, and a model that cannot beat the simple baseline does not ship — which happens more often than the field likes to admit.
Where this appears on the site
We work in PyTorch, and no case study or service page currently published on this site prints it in its stack. Rather than describe an engagement you cannot check, this space stays empty until one does. Ask us and we will talk you through it directly.
Tell us what it is running, what it costs you today, and what you need it to do next. A senior engineer will tell you what we would keep and what we would change.
