Gummi

Agents
Live Stats
Databricks Pipeline Map

About the project

Inspiration

Dexcom's API gives third-party apps glucose data an hour late, on purpose. People with prediabetes or type 2 diabetes (not on insulin) end up with numbers, not coaching. We wanted a coach that reasons about the present, suggests something small and useful before a spike, and is honest about how good its predictions are.

What it does

Gummi is an iPhone CGM coach fronted by a jelly koala whose mood follows your glucose.

Home: coach cards, Gummi, and a chart with Dexcom readings (solid), Gummi's estimate of the delayed hour (dotted), and a 2-hour forecast (dashed).
Chat: "Can I have a cookie?" simulates the food and shows both curves before you eat. "I just had a granola bar" logs it.
Food: a food log where every meal gets a prediction, and 2 hours later Gummi grades itself against two baselines: the published CGM-only model and the last reading.
Activity: cards an event-driven agent posts on its own: morning briefings, walk nudges, meal stories, evening recaps.
Day: time in range, the day hour by hour, and Gummi's report card.
Walks: live steps from the iPhone pedometer, with the walk's effect labeled by source (published research or your own data).

How we built it

Three of us, one lane each:

iOS (Swift, SwiftUI): a procedural 3D koala in RealityKit with jelly physics, rendered at 120 fps through RealityRenderer into Metal. Live updates over server-sent events. Swift Charts, HealthKit, and Core Motion.
Backend (a Databricks App on FastAPI): in-memory live state, the model running in-process, and a chat agent plus four background agents on Databricks Foundation Model APIs. Every run is traced in MLflow.
Data (Databricks): the BIG IDEAs study data (16 participants) loaded into Unity Catalog. A ridge model on CGM history plus meal absorption curves, trained with participant-grouped cross-validation. During the replay, each person is predicted by a fold model that never saw them. A Lakeflow declarative pipeline streams every event into Delta tables, and the gold accuracy tables feed the agents.

Grades use mean absolute error: $\text{MAE}=\frac{1}{n}\sum_{i=1}^{n}\lvert \hat{y}_i-y_i\rvert$.

Challenges

Free Edition's daily compute limit stopped our App, models, and warehouse mid-build. We added a "server asleep" fallback with offline demo data, and the backend now resumes cleanly after an outage.
After the replay passed the end of one participant's data, the estimate kept extrapolating for 65 hours. Gummi now refuses to estimate when the newest reading is more than 3 hours old.
RealityView capped the koala at 60 fps, so we wrote our own render loop to reach 120 Hz.
Replay leakage: one model trained on everyone would grade itself in-sample, so each participant is predicted by a fold model that never saw them.

Accomplishments

Out-of-sample error at 30 minutes (mean absolute error, mg/dL): Gummi 8.9 vs CGM-only 9.5 vs last value 10.7. At 60 minutes: 12.1 vs 13.4 vs 14.9.
Reproduced the published CGM-only baseline (13.90 mg/dL at 30 minutes, RMSE).
In the demo, Gummi predicted 185 for an evening snack, the actual peak was 186, and it beat CGM-only (155).