Language
Code-switching is the norm
Consultations move between Malayalam, English and Malayalam written in Latin characters, often within one sentence.
About
A research initiative, not a product launch.
Clinoria is an independent engineering effort based in Kerala, working on a narrow question: can clinical documentation preserve what a conversation actually meant, in a place where people do not speak one language at a time?
How this began
From a documentation problem to an evidence problem.
The work started with a practical observation: transcription is not the hard part of clinical documentation. The hard part is that a transcript contains claims whose meaning depends on a negation four words earlier, a tense, or which relative the patient was talking about — and that conventional summarisation discards exactly those markers first.
The first build was a radiology-oriented pipeline: extract discrete findings, keep laterality and measurements attached to the structure they describe, render a report deterministically, and refuse to produce a draft the evidence does not support.
Building it surfaced the real problem. The refusals were easy to implement and the extraction was not. Every honest test found another way for the system to be confidently wrong: an invented laterality that passed validation, a Malayalam denial recorded as an assertion, a family history that lost the father and kept the mother.
That is the point at which the project changed character. It stopped being a product with a demo and became a research effort with a defect register.
Operating principles
- Publish failures while they are still open.
- Never publish an accuracy figure without clinician-authored gold data.
- Never show a typed transcript and call it transcription.
- Never use real patient data in public material.
- Never claim a language capability the code does not have.
Where we are
Working prototype. Pre-clinical. No independent validation, no certification, no institutional affiliation claimed. Eight open defects are published.
Why Kerala
The hardest version of the problem is also the most useful one.
It would be easier to build for monolingual English consultations in a market with clean training data. That system would also be less necessary here, and less interesting.
Context
Documentation load is real
Clinicians with high patient volumes carry documentation into their own time. That is the problem worth solving.
Evidence
No local gold standard exists
There is no consented, clinician-annotated corpus of code-switched clinical speech to evaluate against. Building one is the research.
Who is behind Clinoria
An engineering-led effort, deliberately small.
Clinoria is built by an independent engineer working with clinicians rather than for them, based in Kerala. There is no company behind it to announce, no funding round to disclose, and no institutional partner to name — stating otherwise would be invention.
The work is documented in public on this site: the architecture, the defect register, the failure analyses and the retention design. That documentation is the credential we are offering, because it is the only one we can evidence.
When a verified biography, affiliation or funding relationship exists, it will appear here with permission and in specific terms. Until then, this section stays as it is.
What we deliberately do not claim
- A medical background we have not evidenced.
- Academic positions or degrees.
- Hospital or university partnerships.
- Investors, grants or funding.
- Customer adoption or clinical use.
If any of these appear elsewhere in connection with Clinoria, ask us for the source.
Where this is going
Three things, in order.
-
Now
Make the defects legible
- Publish the register and keep it current
- Demonstrate failures honestly, with a person present
- Stop claiming capability the code does not have
-
Next
Get clinical gold data
- Consented, de-identified, clinician-annotated material
- An evaluation protocol reviewed by clinicians, not by us
- A published result that includes what failed
-
Then
Earn the word "validated"
- Held-out evaluation in English and Malayalam
- Independent review of the protocol and the numbers
- No shortcuts: if it does not hold up, it does not get said
Work with us
If this problem matters to you, say so plainly.
Clinicians, researchers and engineers who want to argue with the design are all welcome.