Open any transcript of a real medical conversation in much of the world and you will find something that does not appear in the demonstrations: a sentence that begins in one language, carries a clinical term in English because that is the only word anyone uses for it, and finishes in the first language again. Sometimes a relative answers over the top in a third.
Why this breaks most systems
Many multilingual tools work by detecting a language and then operating in it. That model assumes one language per utterance. When a sentence contains two, detection picks a winner and the rest is treated as noise — so the clinical term, which is usually the most important word present, is the part most likely to be mangled.
The result is a system that performs well in testing, where people speak one language carefully, and poorly in a consultation room, where they do not.
What we do instead
- The message carries its own language, not the account. If someone writes in Urdu today and English tomorrow, each message is handled as what it is. Settings describe a preference, not a constraint.
- Mixed input is expected, not an exception. A clinical term in English inside another language is the normal case.
- The answer comes back in the language of the question. Asking in one language and being answered in another is the product failing at the thing it promises.
Why it matters commercially, not only humanly
The people least well served by existing health information are precisely those who do not read English comfortably. They are also an enormous market that most health AI has effectively ignored, because building for them is harder than building for the languages the benchmarks are written in.
We think that gap is the opportunity, not an afterthought to it.
This article explains how our technology works. It is health information, not medical advice.