Our own AI
In developmentDa Vinci
Da Vinci is made by Cosine. It is the AI inside everything we make. It hears what people say to e1-4, talks back, recognises whose voice is whose, and will eventually reason across all of our sites. It runs on open-weight models on a server we rent and run ourselves, so nothing said to it goes to another company.
What it does today
Hears
LiveTurns speech into text, on the device where it can and on our server where it can't. Whisper large-v3 turbo.
Recognises
LiveTells one voice from another, which is how e1-4 signs people in without a password. ECAPA-TDNN, with the voiceprint sealed under a key only our server holds.
Speaks
In developmentAnswers out loud in its own voice. Kokoro, an 82-million-parameter open model.
Thinks
In developmentCarries a conversation and reasons about what was said. Qwen3, with Qwen3 Embedding for memory and GLiNER2 for picking out names and structure.
How it learns
Plainly, because this is the part people most want to know.
From conversations, as text
What people say to Da Vinci, and what it says back, helps train our own models. This is on unless a person turns it off, which is one tap on their e1-4 profile and also deletes what was kept.
Never from voice
Recordings, voiceprints and voice streams are never training data. The training store has no place to put audio.
Not from everyone, and not everything
Nothing from anyone under 18, or from anyone who hasn't told us their age. A filter drops conversations about health, religion, sexuality, politics, ethnicity, union membership, criminal history or money details before they are kept.
Never for anyone else
We don't sell it, share it, or use it to train models that aren't ours.
What it is told about itself
The Chinese room rule
ResearchDa Vinci is told, in its instructions, that it handles symbols and does not understand them, and a guardrail removes any claim that it understands or feels. Research assistants built on it have to show their working and say how sure they are.