Hello
I am Sambhav, currently i am a student and a developer too, few years back when i saw myself and my friends struggling to focus (most of them were preparing for competitive exams, myself included) i saw 2 things, we were all trying some sort of productivity app but it just didnt work, later i figured out that all productivity apps are inside my phone, the distraction device itself, i also noticed that i would study better if there was some external accountability (like my mom asking about studies or results vs no one really noticing)
Then i went down a rabbit hole figuring out productivity stuff, like everything about it, the research, the psychology, why focus breaks and all that, for about a year i spent my spare time thinking about this and trying to build something that would genuinely increase my study time, i came up with a physical device
So, this device just sits on the desk, no notifications, no feeds, nothing else to do on it, just focus , now it obviously has the basic things like timers (pomodoro/countdown), ambient sounds, simple tasks, etc but then there's my twist, there is a feature where you press a button and say whatever doubt you have and get a precise answer, it could be anything study related, basically quick answers without opening a browser (this was important because that’s where i usually fell into rabbit holes)
there is also a study together feature, its more like commitment, no one can just quit otherwise it affects everyone, and there is voting for breaks too
then we have a progress tracker, i’ve iterated on it a lot so i think its actually pretty solid
there is also an optional system which learns your study pattern over time
a smarter alarm for study sessions (not like a normal alarm)
leaderboard and study rooms
(I really like this one)- optional health system for things like eye breaks (20-20-20), hydration, and posture
On the hardware side its running on a pi zero 2w in kiosk mode so it just boots straight into it and there’s no real escape into other things
well, many apps i made are really good but i would never bother to share it, takes unnecessary effort and i don't really know what works best for me will work like that for others
This is less about “running a 400B model on a phone” and more about clever engineering around constraints.
What’s actually happening is:
in mixture-of-experts only a small subset of weights is active per token
Aggressive quantization
Streaming weights from storage instead of loading everything into RAM
So the effective working set is much smaller than 400B.
That said, the trade-offs are obvious: very low token throughput, high latency, and heavy reliance on storage bandwidth. It’s more of a proof-of-concept than something usable.