T1DMDROID

An Android testbed for biohackers and researchers who run and test machine-learning models on blood glucose. It reads up to four CGMs at once (AiDEX X / LinX, Anytime CT5, Libre 3) over Bluetooth LE and runs your ExecuTorch models on their live output, on the phone. Android 12+, arm64 and x86_64, sideloaded.

Caution

Research use only. It is not a medical device and not clinically validated. Its forecasts and calculators may be wrong and must not be used for medical or dosing decisions, nor to replace a real CGM, its official app, or professional care. No warranty, no liability.

Video Gallery

tall_blood_sugar_graph_that_shows_iob_and_cob.mp4 Readings, the forecast band, and carbohydrate and insulin curves on one graph.

read_from_multiple_sensors_at_once.mp4 Up to four CGMs read at once and recorded side by side.

paint_on_the_blood_sugar_graph.mp4 Freehand drawing on the graph.

drive_on_your_blood_sugar_curve.mp4 Minigames played on the glucose trace.

autoregressive_rolling_for_long_horizon_predictions.mp4 Extends the forecast past the model's horizon by feeding it back in.

see_previous_predictions_with_hindsight_mode.mp4 Scrub through past forecasts against the readings that followed.

test_infilling_capabilities_of_prediction_models.mp4 The model fills gaps in the sensor record.

visualize_model_time_output.mp4 The model's estimate of the hour of day, against the local clock.

advanced_model_evaluation.mp4 Backtests, accuracy, calibration, and Clarke and DTS error grids.

lora_finetuning.mp4 Train LoRA adapters on the phone for forecasting, infilling or backcasting.

log_carbohydrates.mp4 Meals, with a carbohydrate appearance curve you can draw.

log_insulin.mp4 Insulin doses, with an action curve you can draw.

advanced_blood_sugar_statistics.mp4 Time in range, GMI, CV, LBGI/HBGI, ADRR and MAGE.

advanced_settings_panel.mp4 Targets, alarms, dose rails, display and backups.

More

• Alarms: out-of-range and loss-of-signal.
• Nightscout: uploads readings.
• Watch link: pushes readings and forecasts over encrypted BLE to T1DMKDE (desktop) and T1DMAUTO (car head unit).
• Backup: export and restore as one gzipped JSON file.

Your own model

Export an ExecuTorch .pte for the XNNPACK backend and its descriptor, as T1DMAI does, and push both:

adb push my.xnnpack.pte my.descriptor.json /sdcard/Android/data/com.t1dm.app.pub/files/models/

Each descriptor in that folder loads as one model. Its artifact key names the .pte; without it, <id>.xnnpack.pte.

Building

Android SDK 36 and the NDK, JDK 21, and Rust with the aarch64-linux-android and x86_64-linux-android targets and cargo-ndk on PATH.

./gradlew :app:assemblePublicRelease

Running on Xiaomi HyperOS / MIUI

• Settings → Apps → T1DMDROID: Autostart on, battery No restrictions, Pause app activity if unused off.
• System Bluetooth app: battery Unrestricted.
• Recents: lock the app's card.
• Lock-screen glucose: adb shell settings put secure lock_screen_show_silent_notifications 1

Related projects

• T1DMSIM: the simulator whose synthetic traces pretrain the model.
• T1DMAI: trains the model and exports it to ExecuTorch.

License

MIT. See LICENSE.