The problem Watchtower solves

Managing pre-diabetes requires consistent tracking of blood glucose, meals, and medications — but most health apps are bloated, cloud-dependent, and don't connect the dots between what you eat and how your glucose responds. Watchtower is a private, self-hosted web app that lets you log readings, meals, exercise, and medications in under 10 seconds. You can also just type a plain English sentence like "had oatmeal and OJ, bg was 145 after" and Claude parses and stores it automatically. Every night, Claude Sonnet generates a daily digest explaining why your glucose moved the way it did — naming actual foods, not giving generic advice. A live Intel card gives context-aware recommendations throughout the day, and a food insights system automatically learns which of your meals are safe and which spike your glucose over time.

Challenges I ran into

The biggest challenge was making AI analysis actually useful rather than generic. The digest system needed to reference specific foods the user ate (like "the shawarma roomali kept your bedtime flat") rather than giving boilerplate advice like "eat more protein." This required carefully constructing prompts with meal-to-reading temporal links, medication schedules, exercise timing, and next-morning fasting data. Another challenge was the natural language logging — parsing a single sentence into structured meal + reading data with correct types, timestamps, and meal linkages. The recommendation engine also needed smart caching per time-of-day bucket to avoid redundant API calls while still feeling fresh. Finally, keeping the app fully self-hosted with SQLite meant designing migration logic that handles schema evolution gracefully across deploys without data loss.