Mira

Background

Mira is my personal assistant, built on a small local language model with retrieval, speech, and image generation around it. A small model is more fallible than a frontier one, and most of the design goes into making it trustworthy anyway. The model is never the system of record, and where being right matters deterministic code either checks its output or replaces it outright.

Grounded answers

The assistant answers questions through a bounded endpoint with a strict contract. The server stores nothing and the conversation lives only in the client process, with each request carrying the prior turns. Claims about local or current data are rejected until a read-only tool has grounded them, while stable general knowledge may be answered directly. A versioned registry of reviewed facts covers high-stakes topics. Before an answer is delivered, a deterministic validation pass corrects reviewed aliases, generalizes identifiers the sources do not support, and removes any link that neither the registry nor the user supplied.

Reads are allowlisted. The assistant sees only the library roots I have registered, and adding a project to its world means registering another root.

The morning briefing

Each morning Mira assembles a briefing for the menu bar, a notification, and email. No string in it comes from a model. Every line is built from data read out of the registered roots, and a line that changes nothing about the day is cut. Failures stay visible, and a source that cannot be read renders one plain line instead of silently disappearing.

Research runs

The same assistant fronts my scientific calculations. Questions about a run are routed internally to dedicated bundles for results, status, drafting, and cancellation, and cluster jobs are drafted and submitted through their own scripts. Asking what happened to a calculation and asking for the next one to be drafted use the same terminal command as everything else.

Provenance

Mira began in August 2026 and is under active development.