ai · 10 Oct 2026
B-CON AI: four senses, trained for Nepal, hosted by us
Why we replaced every third-party AI service in our products with models we train and serve ourselves, and what that means for patient privacy.
By Administrator
Until this year, parts of B-CON EHR leaned on foreign AI APIs for speech and language. They were good at English and poor at Nepali, and every call sent something about a patient across a border.
We decided that was the wrong trade. Over the past months we built and deployed four models of our own: B-CON Ears for Nepali speech recognition, B-CON Voice for Nepali speech synthesis, B-CON Brain for clinical notes, plans and questions, and B-CON Eyes for reading documents.
What changed for hospitals
- Consultation transcription now understands code-switched Nepali and English, including drug names and 1-0-1 dose patterns.
- Token calling and prescription read-back use a natural Nepali voice.
- Nothing identifiable leaves the hospital network for inference.
What did not change
The doctor still decides. Every draft is reviewed and signed. Drug interactions still come from a curated database, not from a model's memory.
Read more on the B-CON AI page.