A home-grown intelligence
Something in the house is paying attention.
Not an app, not an assistant, not a subscription. A small warm machine on a shelf that holds everything the family owns — and notices what happens.
An ordinary evening
One voice, the whole home
Ask from any room, or tap any screen. It runs the house's music, films and live television, holds the photographs, keeps the calendar, and watches the front door — and answers like someone who actually lives here. Everything else disappears behind it.
Just say it
What it does
Every piece below is running today — on one small machine on a shelf.
Pulled straight out of the air and recorded automatically. It knows which competitions matter in this house, so it simply gets them. No trials, no menus.
Thousands of films, shows and songs — on every screen in the house the instant you want them. Nothing to sign into.
The family's own pictures — a hundred thousand of them — sit on the house's own drives, not on rented ground. No subscription keeps them alive. No company can decide to stop.
Who came, who left, who was carrying a parcel. Not an alarm — a record of the day, the way an attentive person in the house would have noticed it.
Years of private annotation buried in the tags — invisible to every commercial service — turn out to be the fastest way to find a song. And the visualiser from 2001 is still there, still pulsing, because it was wonderful and nobody asked it to stop being wonderful.
“Can't tell” is a real answer here. The still from the door is captioned with the honest time of the last movement — not a comfortable “just now”.
Why it watches the corner
Four of the family's e-bikes have been stolen — not all from the same spot; one went from inside the garage. That is the reason this part exists at all.
Later, an attempt on another one was caught on the doorbell camera, at the corner the house now watches. That footage turned out to matter more than the losses had, because for the first time the household could see the shape of the thing — how somebody moves at that corner when they mean to take a bike, and how plainly different that looks from a neighbour walking past.
So the Hearth was taught it. Not from the internet, and not by quietly improving itself overnight: someone sat down with the real footage, marked what was in it, and then checked the system's judgement against clips whose answers were already known. It learned the pattern from an observed attempt — the way it learns everything else in this house, by being shown.
What it does with that is narrow, and it is honest about being narrow. It watches the corner where the bike is parked and works out, clip by clip, whether the bike is still there. Most departures are just somebody riding off, and it says so — the day's log stops reading Motion and starts reading someone walked past, and the bike left with them. When a departure resembles what it was taught, it says that, and how sure it is. When the picture won't support a call, it answers can't tell instead of guessing.
It doesn't stop anyone, and it was never going to. It notices what you would have noticed yourself, if you'd happened to be standing at the window.
How it learned the family
The house didn't arrive knowing anyone. It was taught — and the teaching is almost comically unglamorous. A screen shows one clip from the front door at a time and asks who that was, whether they were coming or going, whether the bike moved.
Someone in the house sits and answers. A second or two each. The machine guesses first — it puts its own answer up, marked as a guess, with how sure it is — and most of the time the human just confirms it and moves on. The corrections are the point. Each one goes back in. The neighbours start being recognisable as the neighbours rather than as strangers.
You don't get a member of a household off a shelf. You get one by raising it.
How it lives in the house
Say it from the kitchen with your hands full. One listener in the house hears you, and answers out loud in the room.
A panel on the wall, the television, the phone in your pocket — windows onto the same room rather than four separate apps. One thing in front of you at a time.
One small machine on a shelf, joined to a mind that reasons. It learns this household by being told, and stays out of the way otherwise.
No support line
Say what you'd rather it did. That sentence is the mechanism — there is no ticket, no queue, no committee — and often enough it's different by morning.
Are you a techie? Got an HDHomeRun Flex spot-checked in VLC, feeding a Dockerized Jellyfin + Immich stack behind an nginx front door, with Schedules Direct guide data and a Whisper→Piper voice pipeline — all stitched together by an AI that never sleeps?
hdhomerun · vlc · jellyfin · immich · home-assistant · schedules-direct · whisper · piper · docker-compose · nginx · cifs · claude
Yeah. So do we. The difference: nobody in this house will ever see a single pixel of it.
The box
No rack, no server room — one small box on a shelf, warm to the touch, glowing softly, always awake.
It began as a laptop in a closet with the lid shut. It's a grey office desktop now, on a shelf, plugged into a battery so a power cut doesn't take it down mid-evening. The idea hasn't changed.
The unusual part
Most systems like this are designed to always produce an answer, because an answer looks like competence. This one keeps machinery for doubting itself: a real can't tell among the verdicts, captions that give the honest time rather than the flattering one, and a quiet check on a patch of wall that should never change — so when the camera has been knocked or the light has done something strange, the house marks its own conclusions as not to be leaned on.
A system that maintains machinery for doubting itself was built by someone who intends to actually rely on it.
A personality in the house
One that holds what the family owns, knows what's on, keeps an eye on the front door — and tells you honestly when it isn't sure.