GlucoseDAOGet involved ↗GLUCOSE PREDICTION · OPEN RESEARCH · APPS YOU CAN TRY
Connect your own CGM data to GlucoPhone and GlucoWatch, run compatible prediction models on your phone, and help benchmark human forecasting in our Sugar Sugar study.
Connect Dexcom Share or Nightscout
Local model inference · Android + Wear OS
01 / SUGAR SUGAR
Sugar Sugar: a game that is also a study.
Look at three hours of glucose history, then draw your prediction for the next 60 minutes. Compare it with the real sensor readings and get your accuracy score. Your results help researchers measure human forecasting performance against machine learning models.
▶ Play Sugar Sugar ↗For adults 18+, with or without diabetes or a CGM. About 15–20 minutes. Ethics clearance: University Medicine Rostock, A 2026-0064. Read the participant information and give consent on the study website.
02 / THE RESEARCH
Our research includes the SugarOne transformer forecaster and experiments with JEPA encoders for CGM sequences. We compare models across forecast horizons, distinguish familiar from unseen people, and separate past-only forecasts from models with announced future inputs.
A concept illustration of forecasting, using synthetic values.
The numbers below come from our local MetaboNet evaluation of a past-only ensemble. Switch the horizon to see how prediction error changes as the model looks further ahead.
A past-only research ensemble on the MetaboNet sandbox test set. These development results are not clinical validation or a ranking against other teams.
MARD is mean absolute relative difference between forecasts and reference glucose. This row pools 2,648,987 target observations at each horizon and includes both familiar and new people. Model: Blend-EW-4run-CITRASrefitx2-TFTrefitx2-pastonly-s42s45-mv1. Input: past observations only. Four-horizon average: 16.41%. Source: the local research leaderboard reviewed 19 October 2026. Performance depends on the dataset, population and forecast horizon.
REAL DATA · GENERATED BY GLUCOPHONE
GlucoPhone generated this report for the activity on 17 October 2026. It aligns glucose, heart rate, estimated grade-adjusted pace and elevation on the same timeline, with insulin context alongside the activity.
The report distinguishes recorded bolus from basal insulin calculated from rates. Calculated insulin is an estimate, not independently measured delivery.
Actual app screenshots.
Synthetic demo data.
GLUCOPHONE + GLUCOWATCH
Use your own Dexcom Share or Nightscout data in GlucoPhone, our Android glucose dashboard. Import a compatible ONNX model and run forecasts locally on your phone. GlucoWatch brings your readings, trends and phone-generated forecasts to a Wear OS watch face and tiles. The phone also works on its own, and the watch can fetch readings directly.
Android 10+ for GlucoPhone. Wear OS 4+ for GlucoWatch. No project server or Google Play Services dependency. Research preview, not a medical device: keep your official CGM app and alarms, and do not use forecasts for treatment decisions.
GET THE RESEARCH PREVIEW
The phone app works on its own. Add the watch app and face when you want glucose on your wrist.
Use Obtainium to install GlucoPhone from GitHub and check for updates. Select the phone APK from the release.
Set up with Obtainium ↗Add github.com/GlucoseDAO/glucowatch and set the APK filter to ^glucowatch-phone-.*\.apk$. Obtainium installs the phone app; it does not install the watch packages.
Download the Android package, install GlucoPhone, and connect your own Dexcom Share or Nightscout data. Select a compatible prediction model on the phone’s Model tab.
Download GlucoPhone APK ↓v0.1.13 · Android 10+ · About 114 MB. Android may ask you to allow installation from your browser.
All release files & updates ↗Get both watch packages from the same release, then use Wear Installer 2 on your phone to send them to your watch.
Watch installation guide ↗Wear OS 4+. Obtainium manages the phone app only.
OPEN RESEARCH + A STARTUP IN DEVELOPMENT
GlucoseDAO’s public work includes glucose forecasting tools, CGM data conversion, the Sugar Sugar study software, and the Apache-2.0 GlucoPhone and GlucoWatch apps. You can inspect the code, report issues and contribute on GitHub.
Our open-science work was awarded a $50,000 milestone-based BGI Nexus grant in the SingularityNET ecosystem.
Alongside the open-source work, we are developing a startup around glucose prediction. We want to turn the research into useful products and build a sustainable organisation behind them.
We are committed to opening as much as is reasonable as we develop the business. Existing public repositories retain their stated licences. We welcome conversations with research partners, product collaborators and investors.
Discuss the startup on Telegram ↗THE TEAM
Founder · Project lead
Project direction, research partnerships and the Sugar Sugar study. Her experience living with type 1 diabetes helps shape the questions we work on.
LinkedIn ↗
Co-founder · Scientific lead
Machine learning and computational biology. Model architecture, evaluation and the training pipeline.
Co-founder · Technical lead
Software engineering and MLOps. Infrastructure, reproducible pipelines and deployment.
Senior ML engineer
Model development, encoder pretraining and experiments with predictive representations.
LinkedIn ↗
Prof. Irina GaynanovaBiostatistics · University of Michigan
Dr. Renat SergazinovGlucose forecasting · Gluformer & GlucoBench
Prof. Georg FuellenIBIMA · University Medicine Rostock
Dr. Virgil IlianAI adviser
FIND YOUR CONNECTION
Spend 15–20 minutes predicting glucose curves and help build a human forecasting benchmark.
▶ Play Sugar Sugar ↗Install GlucoPhone, add GlucoWatch, report issues, or contribute code to our public repositories.
Explore the repositories ↗Discuss research collaboration, grant support or investment in the startup we are developing.
Contact us on Telegram ↗Join the GlucoseDAO Telegram group to talk with the team.WATCH GLUCOSEDAO
Hear the team explain the project, the science and the ideas behind GlucoseDAO on our YouTube channel.
Yes. Sugar Sugar welcomes adults with and without diabetes or a CGM, and offers public sample data. GlucoPhone and GlucoWatch start with synthetic demo data, so you can explore them without connecting a glucose account.
No. Glucowatch and GlucoPhone are development and research previews, not medical devices. Continue using your official CGM app and alarms, and follow your care team’s guidance.
Explore the BGI-supported open-source work, the research results and the Sugar Sugar study. Contact us in our Telegram group to discuss research, evaluation or support.
