Over four weekends I built a bilingual assistant for my family on local, open-source models. This is the build log — and the design decisions that mattered far more than the model did: a voice nobody wants to mute, a human gate before any action, an agent that finds but never transacts, and privacy enforced in code rather than promised in a policy.

1. I Built an AI Agent for My Family — The flagship — what it does, and why restraint was the hardest feature.
2. A Voice Nobody Wants to Mute — Ground the facts deterministically, then let the model warm the words.
3. Propose, Don’t Act — A human-approval gate between intent and action — and why confidence isn’t correctness.
4. Find, Surface, Hand Off — An agent that surfaces is useful. One that transacts is a liability.
5. A Bot That Teaches — A fact is a broadcast; a question is an invitation.
6. The Homelab Behind the Family AI — The hardware, the model ensemble, and the guardrails — boring where it can be, strict where it matters.
7. Homework Stays Private — Separate the signal from the content: prove it got done without exposing what was hard.

← All series