Alfred, part 3 — hands-free and house-wide
Taking Alfred off the desk and into the house — a trained wake word, phone voice over the LAN, grounded answers with citations, and web search that only fires when I actually ask for it. Still local, still private, still $0.
Problem
Parts 1 and 2 proved the foundation could work — a local voice loop, answers grounded in my own notes. What was still missing was the actual promise of a butler: something I can talk to from anywhere in the house, hands-free, without walking over to one mini PC in one room.
Approach
Trained a custom wake word so Alfred listens for its own name instead of a push-to-talk button, and proved the conversation loop end to end. Got the phone talking to it too — voice working over HTTPS across the LAN, not tied to sitting in front of the desk. Answers stayed grounded in my own notes, with citations and a hard refusal when the notes don’t have the answer — the deterministic-answers rule from part 2 holds. Web search got added as a separate, deliberate capability rather than a default: a self-hosted metasearch that only fires when I explicitly ask for it.
Outcome
The wake word works, the conversation loop is proven, and phone voice is live over the LAN — all of it still local, still private, still $0, all on the one mini PC. Now living with it daily, because that’s the only way to find what a demo never shows: how the always-open mic behaves on false triggers, whether conversation mode survives real household noise instead of a quiet test room, and whether the fluid layout actually holds up on the different screens around the house it’s meant to run on. No verdict yet on any of those three — that’s exactly what daily use is for.
Lessons learnt
Same lesson as the last two parts, holding again: build the smallest real version, then let daily use find what a demo can’t. False triggers, real noise and real screens are exactly the kind of thing that only shows up once something is actually living in the house, not in a test.