How it works

The premise

Record a meeting on your phone and get back a speaker-labelled transcript, a summary, and natural-language search across your whole archive — with no cloud APIs and nothing leaving your network.

The stack is self-hosted end to end: a Rust server with ONNX and GPU inference, a mobile recorder, and a semantic index over the full history. It is built to install with a single command on Windows, Linux and macOS, and carries a stub build so the whole system can be run before GPU inference is set up.

What is included

Speaker-labelled transcription with local LLM summarisation — no cloud APIs

Natural-language semantic search across the entire meeting archive

Mobile recorder app feeding a self-hosted server on your own network

One-command install, self-update, and a stub build that runs without a GPU — built, not yet distributed

How it is built

Scribe is built on Rust, ONNX Runtime, TypeScript and CUDA.

The repository is 53% Rust, 35% TypeScript, 4% PowerShell, 4% Shell by volume — a description of what the work actually consists of rather than what it is filed under.

It carries 61 commits since Jun 2026, most recently changed in Aug 2026.

The method behind it

Every project here starts as a written specification, and the specification stays authoritative once code exists. The repository keeps a build plan with live status, an append-only decision log recording where the implementation departed from the plan, and a handoff document describing the true current state — including the parts that are not finished.

That is why a portfolio this wide is not a shallow one, and it is why this page can describe the product without overstating it.

Current state

In development. Under active work right now.

The source is public. Read it rather than taking this page for it.