My work with Glooko was a year-long capstone project through the UC Berkeley Master of Engineering (MEng) program, done alongside three teammates and an academic advisor. Glooko runs a diabetes-management platform that syncs with more than 200 different insulin pumps and continuous glucose monitors, giving it a detailed view of how patients manage their condition day to day. I led the machine learning work on the team. The scale of the problem is what drew us in. Roughly 80 people are diagnosed with Type 1 diabetes every day in the United States, and the disease accounts for close to $14.9 billion in US healthcare costs each year. Glooko's own platform syncs more than 200 pump and monitor models across roughly 1.5 million users. We worked from a subset of 8,000 of those patients, filtered down to 5,900 after removing incomplete records, spanning profile data, continuous glucose monitor (CGM) readings at 5-minute resolution, blood glucose monitor (BGM) readings at 30-minute resolution, insulin bolus events, logged carbohydrate intake, lifestyle data pulled from Apple Health, Fitbit, and Strava, and app engagement, drawn mostly from a Type 1 cohort. The outcome we optimized against throughout was daily percent time-in-range: the share of a day a patient's glucose stayed between 70 and 180 mg/dL. Type 1 management sits split across two views that rarely talk to each other: a patient managing day to day, and a clinician managing a population. We built four workstreams to work that gap from both sides: predicting time-in-range, clustering patients by management quality, recommending changes drawn from better-managed peers, and forecasting glucose itself. The first predicted daily time-in-range using XGBoost regression, deliberately restricted to pump-usage and pump-settings features and excluding glucose values themselves, so the model had to infer control quality from how a patient used their device rather than from the outcome it was trying to predict. We tuned it with Bayesian hyperparameter search and a cross-validation scheme built around the structure of the data, and it reached about 0.024 mean squared error on validation and 0.040 on the test set. A feature-importance pass on top of that gave us a way to interpret which pump-usage patterns mattered most. The second grouped patients by management quality. We combined t-SNE, K-means, DBSCAN, and ToMATo topological clustering with relational autoencoders, then regressed the autoencoder's latent space against our management metric so patients fell along a continuous axis from well-managed to poorly managed rather than into arbitrary buckets. The third workstream used that axis directly: for a given patient, we found the nearest well-managed patients with similar profiles and suggested a gradient of change toward their centroid, a concrete, personalized target instead of a generic guideline. The fourth, and the one I spent the most time on, forecasted continuous blood glucose itself. We trained an LSTM on a four-hour input window to predict the next one to two hours at 5-minute resolution, validated with a walk-forward scheme so the model was never evaluated on data that preceded its training window. It reached about 32.46 RMSE against a test-set standard deviation of 35.76. We started with 30 LSTMs, one per patient cluster, then used transfer learning to fine-tune each into a model specific to the individual patient, which is what let us scale personalized forecasting without training tens of thousands of models from scratch. The project's stated impact was threefold: better real-time diabetes care through the recommendation engine, customized insights driven by the deep-learning featurization rather than generic rules, and a training architecture built to scale across Glooko's full patient base. We also saw patterns in the data that could plausibly inform how future insulin pumps are designed, connecting the software back to the hardware. The team, Weiling Kang, Jing Wang, Andy Spezzatti, and I, presented the work as a poster titled "Harness the Power of Machine Learning to Help People With Diabetes," under the guidance of our UC Berkeley advisor Paul Grigas and Glooko's own team: Ricardo Abad, Sarine Babikian, and Vikram Singh. Glooko was my first sustained project applying machine learning to patient behavior rather than device telemetry, and it reinforced the same discipline I had carried from Zipline: build the feature representation and the data pipeline carefully before trying to predict anything, and validate a model's output against a concrete, interpretable comparison, a peer's management trajectory or a forecast's RMSE against the data's own variance, rather than a single accuracy number.
