Mosaic: Your Ambient Life Stream

In a post-smartphone world, our devices should fade into the background, not demand constant attention. Mosaic is a personal ambient dashboard that lives on smart glasses, smartwatches, e-ink badges, and smart speakers. It constantly learns your life's patterns and serves up a curated stream of your own data: a photo from a year ago, a note you jotted, a weather correlation with your mood, an upcoming event, a snippet of a conversation. It's like a private, intelligent museum of your life. No alerts, no dopamine loops—just gentle, timely whispers from your past and hints for your future.

Mosaic uses on-device AI to select moments based on context (location, time, recent activity) and emotional resonance. All processing stays local—no cloud, no tracking. The phone is merely a configuration tool; the real magic happens on wearables and ambient displays. You might glance at your watch and see a photo from your last beach trip, or hear a brief note you dictated weeks ago whispered through your earbuds as you walk into a café where you had that conversation. The experience is serendipitous, personal, and entirely private.

The system learns from your feedback (a whisper of 'good' or 'skip') to refine its selection. Over time, it becomes a mirror of your life's texture—a mosaic of moments that makes you pause, smile, and remember.

We have digitized our lives but lost the serendipity. Notifications are either urgent or trivial. We rarely revisit personal data—photos, notes, events—unless we actively search for them. The itch Mosaic scratches is the desire to make your own data delightful and meaningful again, without demanding attention. It's for people who want a personal historian, not a taskmaster. In the ambient-computing era, our devices should enhance our experience of time, not fragment it.

• Deliver a daily 'mosaic' of ~10 items (photos, notes, highlights) via ambient displays.
• Learn user's context (location, time, activity) to pick relevant items.
• On-device processing only; no cloud upload.
• Support for wearable (glasses, watch) and smart speaker output.
• Integrate with calendar, photos, notes, health data.
• Keep user interaction to a minimum—no scrolling, just glance and whisper commands.

• No notifications or alerts.
• No social sharing features.
• No recommendation engine for external content.
• No full-screen phone app; phone is just a configuration tool.
• No storing data on any server.

• On-device ML: CoreML or TensorFlow Lite for scene recognition, sentiment analysis, and clustering.
• Data sources: HealthKit, Calendar, Photos, Notes (via Core Data), Location services.
• Output devices: Apple Watch, AirPods (via dynamic audio), smart displays (via HomePod), custom e-ink dashboard.
• Backend: None; all on-device.
• Prototype: SwiftUI for config app, Swift for ambient agent.

A persistent background service (launchd or WatchOS complication) that periodically evaluates current context and compares it with user's data store. Each data point (photo, note, etc.) is tagged with metadata (time, location, people, mood) and stored as a vector embedding via a lightweight on-device model. A 'relevance engine' scores each potential item based on recency, location match, calendar events, and emotional significance (derived from sentiment analysis of text).

The top-scored items (e.g., top 3) are sent to output devices via local network protocols (AirPlay-like). On watch, it's a glanceable complication showing one item at a time. On audio, it's a short soundscape or whispered thought. User can whisper 'why this' (via Siri or a tap) to get context, or 'skip' to move on. All processing is in background; no user interface except occasional glance.

• Battery drain from constant context evaluation.
• Privacy concerns if user feels 'watched' even locally.
• Serendipity might sometimes feel irrelevant or intrusive.
• Limited adoption due to reliance on Apple ecosystem; Android alternative needed.

• How to define 'emotional significance' algorithmically?
• Should it also integrate with news or current events (ambient news feed) or stay purely personal?
• How to handle deletion of sensitive items without leaving traces?
• What is the right frequency of updates? (currently every 30 min)

The project came before the hardware was ready—ambient displays were still niche, and on-device AI for embedding was too power-hungry. A prototype ran for a week but never felt magical enough. I kept waiting for the perfect smart glasses that never came, and then life got busy. The plan is a reminder of what could be.

This is a design for a future where phones are gone. It intentionally rejects the notification model. The most challenging part is making the selection feel 'smart' without being creepy. I believe a local-only system can achieve that.