I’m building an Android app (Keynotif) to filter notifications while you sleep.
The idea came from a simple problem:
You wake up, see 30–40 notifications, and don’t know if anything actually matters.
Not a volume problem.
An uncertainty problem.
What I have so far
I now have an early version running on my phone.
Current capabilities:
- Listens to all incoming notifications
- Filters obvious noise (system / low-signal)
- Supports user-defined ignore rules
- Respects Do Not Disturb (stops processing completely)
It’s all local-first for now (with a simple DB).
Result:
It does make mornings quieter.
But it doesn’t solve the real problem yet.
What’s still missing
Even with fewer notifications, I still:
- scan everything
- hesitate
- second-guess what matters
So filtering ≠ solving.
What I’m working on next
The hard part is defining “importance”.
Current direction:
-
On-device anonymization
- strip identity
- extract patterns (timing, sender behavior, frequency)
- convert notifications → abstract signals
-
Decision layer (likely server-side)
- learn what I actually respond to
- classify urgent vs routine
- return a simple “this mattered” signal
Open questions
- How do you define “important” without overfitting?
- How much learning before it becomes unpredictable?
- How do you keep it useful but not intrusive?
If you’ve worked on anything similar (notifications, ranking, prioritization), I’d love to hear how you approached it.
Full write-up here:
https://blog.cahyanudien.site/keynotif-part-2-the-part-thats-actually-hard
