Notes from building, not a content calendar.
We'd rather ship than write, so this is a short log — a few honest notes on what we've been working on and why. No dates we don't mean, no metrics we can't back up. It grows when there's something real to say.
Current build recognises eleven Indian languages on-device. Everything else here is marked for what it is.
We say "on-device by default", not "nothing ever leaves"
The absolute version of a privacy claim is the one that gets a company caught. Why our internal rule is one sentence long, and why this post is a direct restatement of it — not a new promise.
Voice dictation sees a lot — here's exactly what betterflo does with it
The floating mic and the keyboard both sit in a text field to do their job. Here's what each actually reads, what it deliberately doesn't, and why password fields are a hard line.
Shrinking the model to fit a budget phone
A good speech model is big. A budget Android phone is not. How quantization gets an accurate Hindi/Hinglish recognizer down to a size that actually loads and runs on modest hardware.
Why the speech model runs on the phone
Getting a Hindi speech model recognising real speech entirely on-device is one claim. Making it actually fit the phones people carry is another. Here's how the pipeline is built to clear that bar — and what still has to prove it.
More notes soon.
Reading about it is fine. Talking to it is better.
Free to start, no account, offline from the first tap — in eleven Indian languages.