Hello all. Marcus (@pacharanero) found me on Reddit and suggested I post here. Thank you for the invitation.
I am a GP in London, in UK general practice since 2008 and working privately now. I have built an AI consultation assistant called OpenConsult. It went public on 10 September under AGPL-3.0. I would like to show it to you, and more usefully I would like to hear what is wrong with it.
What it does
It helps during the consultation, then writes up afterwards.
During: a rough live transcript appears as people speak, and a side panel keeps a working differential, questions worth asking, and signs worth looking for. The panel revises itself as the conversation goes on.
After: the recording is re-transcribed at higher quality and diarised into Doctor and Patient, then a concise SOAP-style note is drafted in which every claim cites the transcript turn it came from. Click a citation and that turn opens. From a signed note, and only from a signed note, it will draft a referral letter. A separate panel answers guideline questions only from a local guideline corpus that the operator supplies, and declines when the corpus does not cover the topic.
Nothing leaves the machine. Speech recognition, clinical reasoning, embeddings and the voice all run locally. The reference machine is one RTX 4090 with 24 GB, and the models take turns on the card. That choreography is the whole trick and also the slowest part.
Every output is a draft. Nothing is final until the doctor edits it and signs it.
That is a whole consultation, five minutes and forty-three seconds, with an acted patient. It has never been run on a real patient and it never will be. It is not a medical device and it has had no regulatory assessment.
There is also an experimental auto mode, off by default, in which the assistant takes the history itself while the doctor supervises. It says plainly that it is a computer, invites the patient to tell their story, then asks the panel’s questions one at a time. The doctor can take it back with one tap. It needs a lot more refinement and I am not claiming otherwise.
The awkward parts, which I would rather say myself
Most of the code was written by an AI. I wrote the specifications, reviewed the work and refereed it. I have not tidied that away. Five runs in a real room turned up four serious defects that the test suite had passed, and each is written up with the reason no test caught it.
An independent security audit found twelve issues, one of them critical. What they were and what was done about them is in the repository.
The project began for Sinhala, because consultations in Sri Lanka are held in Sinhala, code-switched with English medical terms, and written up in English. A pre-registered evaluation showed that no available model transcribes that safely: drug names and numbers did not survive. So version 1 is English-only, and the negative result is published rather than buried.
What I would like from this forum in particular
Integration. OpenConsult has its own small patient database, because it had to start somewhere, and no route into EMIS or SystmOne. I have read Marcus’s guide to GP system integration options here. My honest question is whether a signed note from something like this can realistically reach a GP record through IM1 or GP Connect, or whether copy-paste is the truthful answer for a prototype, and for how long.
Coding the note. The note is prose with citations and carries no SNOMED CT. If you have done that mapping in anger, I would like to know where it hurts.
Where it breaks in a clinic. Everything so far has been a quiet room and a scripted patient. Background noise, a relative in the room, a patient who talks across you, a cheap microphone, a 12 GB card. What falls over first?
What the audit missed.
Archived on Zenodo as v1.0.0, DOI 10.5281/zenodo.22689261
Blunt is welcome. I would rather hear what is wrong with it than what is good about it.