I think in English. Also Hindi. Sometimes Punjabi. Most of the time all three at once — half a sentence in one, half in another. The way a billion of us actually talk.
Every dictation tool I tried flattened that. They all use Whisper, trained on the internet's English. It hears Hinglish and panics.
So I rebuilt dictation on Gemini — which listens to audio natively, instead of forcing it through speech-to-text first.
Gemini doesn't just transcribe what it heard. It understands what you meant. So you can give it instructions — fix my grammar, format as bullets, switch to Hindi. None of that is possible with a transcription model.
It started as a weekend project. My teammates wouldn't give it back. So I cleaned it up, added an API for friends who couldn't access Vertex, and made the whole thing open source.
Now it's yours.