TuckBack

To-do apps optimise the list. TuckBack optimises what it feels like to carry your life. You tuck a thought away (by text, voice, camera or scan, in English, Hindi, Hinglish, Spanish or German) and it leaves your attention. It returns when it can actually move, in your own words, with who it involves and one grounded next step instead of a wall of options.

The AI is designed around one rule: models propose, code decides. Every capture is committed to an on-device SQLite database before any model runs, so an AI failure can never lose a thought. A deterministic interpreter and forty versioned Action Recipes always produce an answer. Apple Foundation Models (iOS 26) can refine it on-device, and an opt-in OpenAI or Gemini second opinion can compete, but every proposal passes the same grounding validator and confidence policy. Ungrounded people, times or amounts, unsafe actions and wrong-language output are rejected. No model can send, pay, book, delete or complete anything.

Learning optimises for loops that actually get resolved, not engagement: evidence builds up gradually, decays after 90 days, and any learned pattern can be forgotten. Repeated snoozes trigger a Rescue engine that offers a smaller step rather than a louder nag. Designed, engineered and shipped solo: native Swift/SwiftUI app, widgets and Apple Watch, plus an optional Fastify backend for accounts, encrypted sync and Daily Return emails.