# Private assistant

An always-on private AI assistant that a small group of people text the way they would text a person, running on open-source models I host myself.

Kind of work: Private AI assistant. Built with: Open-source model, Self-hosted, About 30 tool servers, Text messaging. AI tooling used: Open source, Anthropic.

Status: In daily use.

## The pain

Everyday planning was scattered across a dozen apps and one person’s head, and the people using it wanted replies generated on private hardware I operate, not by a cloud AI service.

Most assistants answer every message from a cloud AI service and charge per message. Some people, and some businesses, want neither.

An assistant that can send, spend or delete needs guardrails that do not depend on the model behaving.

Something people rely on every day has to be watched and improved every day.

## What was built

**Self-hosted.** Every reply is generated on hardware I operate myself, by an open-source model with vision. The cloud chat path is off by default, with spending caps. A daily quality review uses Claude.

**Routing.** Keyword and phrase triggers first, then nearest match on embeddings, then a general-chat fallback. Each turn sees only the tools it needs.

**Everyday jobs.** Planning, lists and reminders, a morning brief, travel times, recommendations, product search, alerts and tracking, all by plain text message.

**Asks before it acts.** A one-time-code confirmation by text before any send, delete or purchase. Text pulled from the web or email can never trigger an action by itself, and identity comes from the messaging layer, not from the model.

**Memory.** Conversations are grouped into sessions; when one ends, the model condenses it into lasting memories. People can view and delete their own.

**Daily review.** A daily quality review proposes fixes, gates them with tests and evaluations, deploys the safe ones and reports back.

## Results

- 32: skills.
- About 150: tools across 30 tool servers.
- More than 7,000: automated tests.
- About 1,200: evaluation cases across a dozen suites; 60 drawn from real use.

## How it was built

- A model swap is judged by a decision rule fixed before the benchmark runs.
- A reply flagged as bad is turned into a test case by the daily review.
- Staged deploys: one script, a lock, a health check and automatic rollback.
- A ledger of fixes, with trend metrics against a seven-day median.
- A health command that reports each component and never prints secrets.

## What it means for a business

A small business can have an assistant that works inside its own tools, keeps its data in-house, asks before it acts, and is monitored and improved every day.

[Work with me](https://lichucki.ai/contact)
