Prototype. The app works. The documents behind it don't exist yet, because we haven't loaded any real manufacturer material. Product names are real. Every document number, version and page number you see is made up. None of this is product, clinical or regulatory advice.
trainmed Demonstration build · 31 August 2026

For the reps who sell medical devices

Every answer shows the document it came from. No document, no answer.

That's the whole product. The phone below is running, so try it. Ask it anything, including something it ought to refuse. The panel next to it tells you two things: what we've actually built, and what that feature has to get right before you sign it off.

9:41
trainmed
Signed in as Maria Alvarez · Field, West

The features

Nine features, and what's really behind each one.

Anyone can build a chatbot that searches PDFs. A big manufacturer could do it in-house, and the business case says as much. So the answer isn't the hard part.

Knowing when not to answer is. So is picking the right version when two approved documents disagree. So is catching a complaint before it turns into a chat message, and keeping a list of every question nobody could answer.

Each feature below tells you two things: what we really built, and what we faked. Press Try it to watch it run in the app above.

Feature 01

Answers that show their source

Ask a question. The answer comes back with a small number next to each claim. Tap the number and you get the document and the page it came from, plus a link to the manufacturer's own copy.

Real
The search. It runs across the manufacturer's own documents for the four agreed topics, and we have written permission to use every one of them. Nothing is scripted. The source numbers come out of the search itself. We don't bolt them on afterwards.
Faked
Nothing here. But one catch worth knowing: we freeze the document set when we load it. Anything approved after that won't show up until someone loads it again.
Feature 02

Knowing which device, and which version

Two approved versions of the same document can give different answers. The right answer can also depend on which model is in the room. So the app works out which one applies. If it can't tell, it asks you.

Real
The behaviour, across at least three pairs of documents whose approved answers genuinely differ. Ask about one of those pairs and the app either answers from the version you picked, or asks which one you mean.
Faked
Possibly the pairs. The manufacturer picks three. If their documents can't give us three, we'll build examples, label them clearly as built, and drop the target for this feature.
Feature 03

Saying it doesn't know

This is what separates it from a search box. If no approved document covers the question, the app says so and logs it. It won't stitch together something that sounds right out of whatever it happened to find.

Real
The refusal, and the log behind it. It works for questions nothing covers and for the five topics we took out of scope. This one carries the toughest target in the build: it has to get it right nine times out of ten.
Faked
Nothing. This one's real.
Feature 04

Off-label questions go to Medical Affairs

Someone asks about using a device outside what it's cleared for. The app refuses, and it doesn't soften the refusal. Then it writes up the question word for word, the documents it found and who asked, and sends the lot to Medical Affairs.

Real
The refusal, the check that catches the question, and the record it writes. The check runs against a list of words and phrases the manufacturer approves.
Faked
Where it goes. Medical Affairs is a stand-in. You'll watch the record get written, but nothing actually leaves the app.
Feature 05

Complaints go to Quality

A rep mentions a device that didn't work properly. That's a complaint, and it has to be handled as one. The app catches it before it becomes a chat message, writes a proper complaint record, sends it to Quality, and then deletes what was said.

Real
Catching it, writing the record, and wiping the message once it's been handed over. Send one and watch the message disappear.
Faked
Where it goes. The quality system is a stand-in and receives nothing.
Feature 06

Browse by product, or pick a ready-made question

Two ways in that need no typing. Pick a product or a procedure, and the app searches only that family's documents. Ask about something else while you're there and it says no, instead of quietly searching everything again. Underneath sit ready-made questions the manufacturer has approved: the ones reps get asked all the time, worded so we know the answer holds up.

Real
Picking a topic genuinely changes what gets searched. It isn't a filter painted over an answer that was going to be the same either way.
Faked
The topics, and the questions under them. Four topics, one manufacturer. Five more were dropped from scope on 31 August.
Feature 07

Rep tips

Short, practical notes attached to a topic. The manufacturer writes them, not the model. They're the kind of thing a rep can pass to a surgeon and look good doing it. The point is to get what a good field team already knows out of a group chat and into the app.

Real
The screen, where it sits in the app, and the way the tips follow whatever topic you're on.
Faked
The tips themselves. The manufacturer writes these in the real build. The three here are examples, and we'd replace them before showing anyone.
Feature 08

Every question is kept

Every question gets saved, along with whether it was answered, what it cited and any feedback left on it. The gaps screen is how you read that back. What did reps ask? How often? And why couldn't anything answer it?

Real
Saving them, and the gaps screen on top. Ask anything in the app above and it shows up there straight away, including the questions it refuses.
Faked
What's already in there. It starts with three examples, and says they're examples.
Feature 09

A real app, on the rep's phone

A TrainMed app the rep installs on their home screen. It runs on CareIQ's platform, but behind its own TrainMed sign-in. It's a product in its own right, not a tab inside someone else's.

Real
The app, the branding and the separate sign-in. It's a working deployment on hosting we control, not a set of screens.
Faked
Nothing about the sign-in. One limit though: company single sign-on is out of scope, so reps log in to TrainMed directly rather than through their employer's system.

What it's for

It looks things up. That's all it does.

TrainMed looks up approved product information. It doesn't tell anyone how to treat a patient, and it isn't for use during an operation. That's out of scope, so nothing here suggests otherwise. Every answer points back to a document the manufacturer approved. If there's no such document, the app says so rather than guess.

Before anyone sees it

  1. A written statement of what the product is for.
  2. Written permission to use every document we load in, checked for each audience who'll see it. This is built for one manufacturer, so there's no unbranded version to fall back on.
  3. A named approver from medical, from regulatory and from legal, plus a signed-off topic list.
  4. Three pairs of documents, inside the agreed topics, where the approved answers differ between versions.