Rounds

Our first version could hold a convincing conversation, but it could not reliably hold the clinical truth. Ask the same question twice and part of the history might shift. Request an undefined test and it could invent a plausible result.

We rebuilt the boundary. The language model now handles the conversation, while the patient's facts, investigation results and encounter evidence stay outside it.

One case is open if you want to try breaking it:

https://roundsclinical.com/simulator?guest=1&utm_source=indiehackers&utm_medium=product_update&utm_campaign=aug11_build_story