The problem
There's so much AI help for productivity now, and almost none of it solves the actual hard part: keeping yourself accountable. A model can plan your week in seconds — following through is the thing people quietly fail at, alone. That's why Cybond has a social layer on top of the tracking. Your schedule can be visible to people you choose, and being seen changes what you do.
The approach
The core model took real iteration to get right. An activity is a metadata-only template — a title, session length bounds, a weekly target — with no schedule of its own. An event series holds a recurrence rule and materializes independent events on a rolling eight-week window; each generated event is an ordinary, fully editable calendar row with no sync back to the rule. Progress is a weekly fit score, completed sessions over target, derived on read.
On top of that sits Cy, the AI assistant. It reads your activities and week, and proposes calendar changes as approval cards — a dashed preview of the event on that day's schedule, with a confirm and a deny. Nothing writes to your calendar without your click, and a deny sends your reason back so Cy revises. The social layer gives every user a public profile with a read-only week view, governed by per-event visibility: public shows details, busy shows only the time block, private never leaves the server.
Google Calendar is opt-in through incremental OAuth — sign-in asks for identity only, and connecting your calendar is a separate consent for exactly two scopes. Stored tokens are encrypted at rest with AES-256-GCM.
What I'd do differently
Start with the boring data model. The first version had one polymorphic table for everything — tasks, goals, habits — with recurrence groups and tombstone rows to keep deleted occurrences deleted. Every feature paid a tax to that flexibility. The rebuild went the opposite way: plain independent event rows, recurrence as a generator with a high-water mark instead of tombstones, and templates that carry no schedule at all. Fewer concepts, and each one does one thing.
What broke
All-day events vanished. The event dialog stored them with a null start time, and the calendar filters rows without a start off the grid — so creating an all-day event made it unrecoverable through the UI, with every unit test green, because the tests mocked the calendar and never rendered the consequence. It only surfaced in hands-on QA. All-day events now persist real midnight-to-23:59 instants plus a flag, and the lesson stuck: for interactive UI, mocked tests prove you passed props, not that the feature works.