Picaku
An AI meeting assistant. Record a meeting, get its summary, decisions and action items, then chat with everything you have ever discussed.
Chunks from six meetings in one vector store. A scoped thread filters before it searches. Illustrative, from the documented RAG design.
The sync endpoint saves the raw transcript and returns 200 at once. Extraction and embedding run behind it, so a client never waits on a model.
Division of labour
No single model runs the pipeline. Whisper transcribes. A reasoning model, routed through OpenRouter, turns the transcript into strict JSON — title, a summary scaled to the meeting’s length, decisions, action items — and answers chat.
Embeddings run inside the backend on ONNX Runtime: 384 dimensions, no paid embedding API. Torch was removed from the production image entirely, and the new vectors were parity-checked against the old ones at cosine 1.0.
Why the model stays in the cloud
Elsewhere on this site a phone runs the model. This one wrote down why it can’t: a model large enough to return valid structured JSON every time does not fit in a phone’s memory once the OS has taken its share, and one that fits is not reliable enough. The transcript leaves the device for sync and embedding anyway.
Scope belongs to the thread
A note has one long-lived thread, and its scope is fixed when the thread is created. A later message that omits the note id inherits it — a client resuming a thread by id alone used to silently widen it to every note.
Short-term memory is the last ten messages, read from the database rather than from the history a client uploads. Both clients had been echoing the question back, so the model saw it twice.
Drafts, never sends
A LangGraph agent writes the follow-up email and saves it as a Gmail draft. It runs only when asked — drafting every meeting automatically flooded inboxes — checks that the meeting has real content first, and never sends anything.
Status, honestly
In production on FastAPI, Supabase Postgres with pgvector, and Google sign-in. A Flutter client is in progress against on-device speech models served from Picaku’s own CDN.
The repository is private, so there is no source link here.
Picaku
Valorant esports intelligence — live match, team and player data from VLR.gg, paired with CYPHER, an analyst grounded in it.
continue