SAYPO is a fitness platform: a Capacitor + React mobile app on iOS and Android, an AI coach running on Gemini, offline-first SQLite for the workouts that have to survive a dead gym Wi-Fi, and a NestJS backend of 98 endpoints behind all of it. It validated to roughly $62k, and most of what made that possible was boring, deliberate scoping.
Offline-first is a data contract, not a feature flag
The app has to log a set, a rep count, GPS distance and step data whether or not the phone has signal. That means the local SQLite schema is the actual source of truth on-device, and sync is a reconciliation job, not a save button. Getting this wrong looks like duplicate workouts after a sync — getting it right means writing the conflict resolution rule once, in one place, before a single screen calls it.
Where the AI coach actually sits
The Gemini-powered coach does not have write access to the workout log directly. It reads the same structured history every screen reads, proposes a plan, and the plan is written back through the same validated endpoints a manual edit would use. That boundary is what keeps a model’s occasional bad suggestion from becoming corrupted user data — the AI is a client of the API, not a bypass around it.
98 endpoints, one ops dashboard
The backend grew to 98 endpoints because a real product has real edges: subscriptions through RevenueCat, health data sync, coaching history, device state. The 20-tab ops dashboard exists because a system that size is unmanageable through direct database access — every one of those tabs maps to an endpoint group, not a table, so the admin surface and the API surface can never silently drift apart.