BOLUS

I built BOLUS because one of the hardest parts of living with Type 1 diabetes is remembering one small action in the middle of an ordinary day: taking a meal bolus. Between classes or at lunch with friends, it’s easy to forget whether I already handled it. My pump can deliver background insulin and automatic corrections, so seeing insulin on my timeline doesn’t settle the question. BOLUS brings Dexcom readings, pump events and meals onto one chart. It distinguishes an automatic correction from a meal bolus and lets me confirm what happened without guessing that I missed a dose. GlucoseCast addresses a question a glucose reading alone can’t answer: where might this trend go over the next hour? It uses recent glucose history and time of day to estimate glucose 15, 30, 45 and 60 minutes ahead. The chart separates measured values from the forecast and shows an uncertainty band. When readings are stale, changing too rapidly or outside the conditions the model can support, GlucoseCast withholds the estimate. We evaluate its forecasts against a simple baseline: what if glucose stayed at its latest reading? Pump history appears on the timeline for context; the current model uses glucose history alone. MealCast tackles the missing meal record from two directions. First, it uses setup routines and the timestamps of confirmed meals to estimate usual eating windows, labeling whether a window comes from the person’s routine or their logs. Second, it checks fresh sensor readings for a sustained, sharp rise, using the size, speed and duration of the change. A qualifying rise prompts “Did you eat around this time?” Only confirmation creates a meal record. An unanswered prompt stays unknown. Neither a glucose spike nor an automatic correction proves someone ate. Meta’s Muse gives everyday language a role in this workflow. For “I might eat after my next class,” Muse identifies the relevant schedule entry or asks which class the person means. BOLUS checks that interpretation against the available records and leaves the plan tentative until confirmed. It offers no insulin-dose or treatment recommendations. The public repository includes an interactive fictional showcase with scripted glucose forecasts and a separate, optional Muse integration reviewers can run with their own key. The connected app’s forecasts are experimental; their accuracy for an individual remains unproven.