TMDream, clinical decision support

This is the case where what I argue for is clearest: a domain where being wrong has consequences, data that cannot leave the organisation, and professionals who only adopt a tool if they can check why it says what it says.

What it does

• A clinical reasoning engine with LLM and RAG, running an open model self-hosted on a private GPU: clinical data never leaves the infrastructure.
• A conversational clinical interview in three languages, reports for clinician and patient, telemedicine with its own scheduling and payments.
• Quality watched by a regression suite and periodic adversarial audits of the reasoning: false positives, false negatives and red flags.

Why self-hosted

With clinical data this is not an aesthetic preference. A medical record travelling to a third-party API is a compliance problem before it is a technical one, and once you frame it that way the conversation with the hospital changes. The model runs on a private GPU and the only thing that leaves is the answer.

What that really costs, without decoration, is written up in Spanish: lo que cambia al autoalojar un modelo.

How we know it is right

A suite of nearly 300 clinical cases with the correct answer recorded, run in full on every change. No prompt, catalogue or model change ships without passing it. And every failure found in real use enters as a new case before it is fixed, so the suite grows by itself and in the right direction.

That is also where an uncomfortable lesson about confidence scores came from, which I wrote up in the 92% that measured nothing: it was my own system.

Pieces I can build on their own

All of this was built inside TMDream and works just as well outside it:

• Scheduling and video consultations. Booking from the website, reminders and a video room created automatically for each visit. It replaced a paid external tool that also took the data outside.
• Online payments that do not slip through. The appointment is confirmed when the payment provider confirms the charge, not when the browser returns to the site. It sounds like a detail and it is the difference between getting paid and giving visits away.
• Automatic PDF reports. Each case produces its document, one for the clinician and one for the patient, with the same design and with nobody copying and pasting anything.
• Admin panel. Schedules, users and permissions: view, edit, move or cancel appointments with automatic notices to whoever needs them.
• Sign in with Google. Log in and register with an account you already have, with manual approval for new sign-ups so not just anyone gets in.
• An installable mobile app, with no store. The platform also works as an app installed from the browser: it goes on the home screen and behaves like an application, with no App Store, no reviews and immediate updates.

If any of this sounds like what you need, tell me about it and I will tell you what I would do, how long it would take, and whether you actually need AI to solve it.