Small models for one enterprise task
A tool that turns a spreadsheet of labelled text into a small sorting model that runs on your own computer, says when it is unsure, and improves as you correct it.
Many enterprise AI jobs are narrow: route a support ticket, tag a request. A small model trained for that one task can be cheap, private and fast. On the public Banking77 dataset (customer-service questions, 77 categories), a model built through the tool scored 92.9% on the 3,080-question test set, on a laptop CPU with no GPU, at about 7 milliseconds per request.
The repository is private, so the code is not linked here. The research behind it is.