Inspiration

Olivia's uncle has diabetic retinopathy. One day she realized he had been missing a step in his insulin routine: he couldn't see that his pen needle has two caps, and only one was coming off. That meant his insulin may not have been going in at all, and there was no way for him to tell from the outside. It stuck with her how easy it is to get an injection wrong when something gets in the way, and how invisible the mistake can be.

Diabetes care is also changing fast. Glucose sensors (CGMs) are now sold over the counter, and GLP-1 medicines like semaglutide and tirzepatide have gone mainstream, often alongside insulin. That means more data and more possible reasons for a high or low reading. A CGM can tell you your glucose is high. It can't tell you whether your insulin went in, whether your GLP-1 is changing how your meal hits, or whether it's something else.

We first wanted to build a tool for people with vision loss. But fully supporting someone who is blind would be very hard, and they would still need a person to help. So we pivoted to people newly diagnosed with type 2 diabetes who are starting insulin, to help them ease into their routine. Let's be honest: nobody knows how to stick themselves with a needle the first time, and with the rise of self-injectables we wanted to build a tool that could help with that.

What it does

Prime is calm, step-by-step help for people on insulin, using their own CGM readings and dose records:

Pen check: speaks the steps for their exact pen, one at a time, with Back / Repeat / Next. It includes the step that's easy to miss: both needle caps come off.
Camera check: they hold up their CGM sensor kit and Prime takes the photo by itself (hands-free, no tap) and checks it.
Ask why: they can talk or type questions like "Why were my mornings so high?" and get answers built from their own CGM readings, insulin and GLP-1 doses, and meals, with sources you can check.
Visit note: a one-page summary for their care team (time in range, overnight lows, dose changes).
Private by design: names are replaced before anything is sent, and every model call is listed on a Privacy screen.

How we built it

Prime runs in a container on a Nebius Serverless Endpoint, using NVIDIA open models through Nebius Token Factory:

Nemotron 3 Super answers questions
Nemotron 3.5 Super VL checks camera pictures
Nemotron 3 Ultra writes the visit note

Records, CGM readings, and camera pictures stay on the device until a question needs them. Code computes every number (time in range, highs, lows) and removes any dose advice before the models write. The API key stays on the server.

Challenges we ran into

Vision works on still frames, not live video. The vision model can't follow a live camera feed. It can only look at single frames. So Prime had to decide on its own when to take the photo: it waits until the user holds the kit still, then snaps a frame and sends it to be checked.
Picking the right model for each job. We had to work out which model does each task best. We ended up giving each job its own Nemotron model: Nemotron 3 Super for answering questions, Nemotron 3.5 Super VL for camera pictures, and Nemotron 3 Ultra for writing the visit note.

Accomplishments that we're proud of

A working app (the demo is a real recording with a fictional patient) using three Nemotron models end to end on Nebius
A hands-free camera check that confirms the right supplies before a dose
Privacy you can see: every call is listed, and names are replaced before sending

What we learned

Scope matters. We started out wanting to help people with vision loss, but we learned that some problems still need a person in the loop. Narrowing to new type 2 diabetes patients starting insulin let us build something that actually works.
A glucose number doesn't explain itself. With CGMs and GLP-1s both common, the same high reading can have many causes. We had code compute the numbers and let the model explain them, citing sources instead of giving advice.
Design around what the model can't do. Vision models look at still frames, not live video. Instead of fighting that, we made the app wait for the user to hold still and take the photo itself, which also made the experience hands-free.
One model doesn't fit every job. Matching each task to its own Nemotron model (questions, camera pictures, visit note) worked better than forcing one model to do everything.
Trust comes from showing your work. For a health app, listing every model call and replacing names before anything is sent matters as much as the features.

What's next for Prime

Heavy testing. Test Prime again and again, especially to make sure the models don't hallucinate.
Real CGM connections. Connect to real sensors instead of a fictional patient's data.
More robust. Harden every step so Prime keeps working reliably for people who depend on it.
Live-streaming vision. Move from still frames to live video, so Prime can watch and guide the user in real time instead of waiting for a single photo.