What will that meal do to your blood sugar?

Show us what's on your plate and we'll sketch your next three hours. Spikes, dips, the lot!

Open the app

Runs right in your browser, nothing to install. The first load takes a minute or two.

One real day of someone's glucose, as the sensor saw it. The yellow dots are meals.

Same bowl of pasta, different person, totally different spike. So we learn you first.

Glucklich predicting Maya's lunch: the dashed prediction, its band, and the real readings in green inside it

Maya's lunch. Dashed: our prediction. Shaded: where 8 in 10 readings should land. Green: what really happened.

What you can do

Predict a meal

Drop your glucose history and a meal, typed or from a photo. Get the next three hours with an honest band around it.

Predict a whole day

Plan meals, exercise and insulin, and see the day's peaks, lows and time in range.

Check us

Drop what really happened, one file or a whole folder. We score the prediction and learn from it.

How it works

The site's How it works page walks through every step on one real person, Sam, with his own data and charts. In short:

Raw sensor data

Remove noise

Type 1 or not?

Photo to nutrition (CLIP + USDA)

Food and insulin in the blood (OpenAPS)

Everyone, people like you, you

Your curve and its band

Read the plate

CLIP, an open image model, gives a phrase for the photo. The best phrases are matched to 5,200 USDA foods at a typical portion, and to meals you logged before that look the same.

Learn you first

A regression on your recent glucose, food, insulin and activity, fitted in three layers: everyone, people like you, then you. Your own data counts most once there is enough of it.

Say how sure we are

We replay your last 20 meals and size the band so about 8 in 10 of your real readings fall inside it.

More: cleaning and classification, how food and insulin reach the blood, the second model, and how we validate. Summary in the README; the full walk-through is in the app.

Results

On people the model never saw: learn from days 1 to 7, predict every meal from day 8 on, scored against the raw sensor.

	Typical miss (RMSE, mg/dL)	Truth inside the 80 % band
Healthy, prediabetes, type 2	22.4	80 %
Type 1	41.6	82 %

RMSE: the typical distance between the real readings and our predicted line. Type 1 is harder: logged carbs are rough, and insulin pumps react to glucose as it changes.

Try it

The whole site runs in your browser: the demo people, Meal, Full Day, Log and How it works. Python and the models load onto your device the first time (a minute or two), then everything runs locally.

Open the app See the code

Or run it on your own machine (faster, and reads food photos):

git lfs install
git clone https://github.com/AlmostAPhysicist/HealthHack2026
cd HealthHack2026
pip install -r requirements.txt
streamlit run app.py