Avyli
A voice-first personal assistant for iOS. Say what you need to do, and it figures out which tasks you are likely to put off, then breaks them into a first step you can start right away.
Avyli
Say it once. Let it handle the rest.
Avyli is a voice-first personal assistant for iOS. You describe your day the way you would to a friend ("English class at 3 tomorrow, go for a run in the evening, and pick up a parcel") and Avyli turns it into structured items on your calendar, reminders, and notifications.
The Idea: Friction
Most to-do apps treat every task the same way: a title, a time, and a reminder. But having time for something doesn't mean it will happen.
"Meet my coach in the evening" happens on its own, because someone else is waiting. "Go for a run in the evening" depends entirely on me getting started, so it usually slides to tomorrow. The first only needs a reminder. The second needs to be broken down.
Avyli is built around that difference, which I call friction. Everything it captures falls into one of three kinds:
| Kind | Meaning | Examples |
|---|---|---|
event |
Happens at a specific time and place, often with others | A class, a doctor's visit |
errand |
A concrete chore that takes minutes, not hours | Buying lunch, a parcel |
task |
Needs sustained effort and produces a result | Problem sets, a report |
Only high-friction items go through decomposition. Avyli doesn't just remind you to "write the report"; it gives you a first step small enough to start right now.
How It Works
- Voice capture with on-device speech recognition
- Two models for two jobs: parsing needs consistent structured output, conversation needs to understand context, so they are handled separately
- System integration through EventKit, widgets, Control Center, Siri, Shortcuts, and Spotlight
- Sync with SwiftData and CloudKit
- A small backend on Cloudflare Workers, protected end to end with App Attest
- Evaluations for model behaviour, covering capture, decomposition, chat, and memory
Status
Avyli is still in development and the source is private. I'll share more when it is ready.