Industries
Healthcare
We turn the appointment, lab, consultation and support data piling up inside hospital software into forecasting and language models embedded in that software.
A hospital’s software produces a large number of the same kinds of records every day: appointments, lab results, consultation notes, support requests. Most of them are kept only to get that day’s work done. Yet which patient will miss their appointment, which request belongs to which unit, or what from a consultation should go into the record can all be learned from the history of those same records.
The model runs inside the software
In healthcare a model’s value shows up not as a separate report but inside the screen the user already has open. The person planning appointments should see the no-show probability on the appointment screen, not in a separate report. The answer to a hospital’s request, and the unit it belongs to, should be ready the moment the request lands in Jira. So we start by deciding on which screen, and at which decision point, the model will be used.
The data stays in the institution
With patient data, where the data goes matters as much as how good the model is. For language models we use two routes side by side: large models through APIs for quick experiments, and fine-tuned local models for work where the data has to stay inside the institution.
Live first, then R&D
We split the ideas on the table into two groups. Work whose data is ready and whose result can be measured goes live quickly. Less certain work such as image processing or conversation analysis is first built as a prototype, and moves into the product only if the results are good enough.