10 Things It Does, and How to Build One
One computer in a closet now streams our home videos, answers the kids by voice, turns grandma’s stories into cartoons, hosts family trivia, and runs a private research engine for work, with most of it staying inside our house.
▶
💬
✨
🖼
🏠
🛡
TheWith an RTX 5090 (32 GB) and a few free, self-hostable tools, one machine in your home runs the AI models, streams the media, indexes the photos, and renders the cartoons.
The goal isn’t to build a smaller ChatGPT in my house. It’s to build an AI that understands my house.
— Albert RicherEvery family I know is renting its own memory. The photos live with one company, the videos with another, the homework helper charges by the question, and the documents are scattered across three clouds. I got tired of it and built the alternative: a single computer in our closet, built around an ASUS TUF Gaming RTX 5090 OC with 32 GB of memory, that does all of that and a list of things none of those services would ever do, because none of them know that the dog ate the birthday cake. I am not a company. I am a dad who did this over a stretch of weekends.
Every chapter carries a rating, a setup time, and a run time, so a technical reader knows what to expect and a curious dad knows where to start.
Setup is my wall-clock time to get it working, after the box itself existed. Each run is how long the machine takes to finish one job. Your part is the human minutes involved. The smaller chips say local or not, web or not, and how hard it leans on the graphics card.
01 · THE MEMORY BANK
The phone hears the request. The closet answers with the video.
Every photo and home video the family has ever taken is indexed by face, by place, by date, by the words visible in the frame, and by what is actually happening in it. That last one is the difference from a folder on a hard drive. These are real searches from our box:
The voice part is a named shortcut, not a replacement for Siri’s brain: say the shortcut’s name, then the request, and it sends the words to the box over our private network and shows what comes back.
The face index took one night to build across our whole library on the 5090 and has run in the background ever since. Immich’s own docs say plainly that its database backup is not a backup of the photos. Both need copies. I learned that before I needed to, which is the only good way.
02 · THE FAMILY ARCHIVE
Record an hour. Keep it forever. Then let the kids watch it.
Sit a grandparent down with a microphone and ask about 1962. The box transcribes every word, pulls each story out as its own chapter, and files it under the people and places it mentions. Then it hands the story to the cartoon studio, and the kids watch grandpa’s first car break down on the way to the dance, drawn in the family’s own style. Her recipe cards go in the same drawer: scanned, read aloud by her, filed with the story of the first time she made each one.
I recorded the first interview on a phone at the kitchen table. Whisper produced the transcript; the language model then corrected every family name against the list of people in our photo index, which is why none of them came out wrong. The kids have watched the first illustrated story more times than anything on a streaming service.
03 · THE BEDTIME ENGINE
A series, not a story. Chapter 12 picks up where chapter 11 left off.
Her button asks three things: who is in it, where are we going, and how long. The box knows her name, her best friend’s name, the dog, the tree house, and the dragon she invented in March. Each night it writes the next chapter in that world, at her reading level, with a cliffhanger she will ask about at breakfast, and reads it aloud in a warm local voice. On nights a parent is traveling, the chapter can be read in that parent’s cloned voice, from a phone, from anywhere, if that parent recorded the samples and said yes.
The story bible started as a note with her name, the dog, and the tree house. It is now pages long because she edits it. My rule: a chapter is never longer than the time it takes her to fall asleep, so the model writes to a word count I tuned by ear.
04 · THE CARTOON STUDIO
Your family, drawn once, starring forever. Same people, new adventures.
Each of us is drawn once as a simple cut-out character, the way a certain Colorado town has been drawn for thirty years. So is the house, the car, the dog, and the kids’ friends. After that, an episode starts as a sentence at dinner. Something ridiculous happens with the spaghetti, someone says “Family AI, cartoon that,” and by bedtime there is a new episode under Family TV called The Great Spaghetti Disaster. Theoretical ones work too: “What if our family went to the moon?” Funny things that happened, things that might, and grandma’s stories from chapter 02 all run through the same studio.
Here is the part I got wrong the first time. I assumed a script goes into a video model and an episode comes out. It does not, or rather it does and Dad has a different face in shot 17 and the dog has five legs. The pipeline that actually works is less generative, not more: fixed 2D rigs with reusable heads, mouths, arms, and poses; a language model writing structured scenes; voices generated separately; a lip-sync tool timing the mouths; the rigs animated programmatically; generative video only for the occasional special shot. Every shot is eight to fifteen seconds, then they are stitched.
The characters were the first thing my kids asked for and the reason they stopped thinking of the closet computer as Dad’s. Drawing each of us once, then locking the style, was the whole trick. The pinboard in the picture is a fair drawing of the real one. It is the most fun thing on the box and the least finished, and both of those are true.
05 · THE WHY BUTTON
One button on the playroom iPad. Answers pitched to whichever kid it belongs to.
The app has a single red question mark. Press it, the screen says “Asking the All-Knowing Daddy,” and the child asks whatever is on her mind. The box answers in a way that particular kid will understand quickly: shorter and more concrete for the seven-year-old, more detail for the ten-year-old. If it is a good one, it offers a five-minute cartoon on the subject.
The first version recognized the child by camera. I took that out. The playroom iPad is simply logged into the child’s own account, which solves the same problem without biometrics, and that account has no work documents, no household finances, no unrestricted tools, and no web at all. The kids’ AI is fun without holding the master key to the house.
The name came from my daughter. I wrote each kid’s tone rules as plain sentences and I read the question log on Sundays. The best part is watching a seven-year-old learn that “why” is a button she is allowed to press as often as she wants.
06 · THE HOUSE MANUAL
Photograph every label once. Ask forever. Then do the attic.
Walk the house and photograph every appliance label, model plate, warranty card, breaker panel, and paint can. From then on the house has a manual that answers questions: which breaker is the garage, what size is the furnace filter and when is it due, what color is the hallway, when was the water heater installed. Receipts go in the same pile, so finding the invoice from the HVAC replacement becomes a query. It is a document brain, not a tax preparer.
Then keep walking, into the attic and the garage. Photograph the bins, the shelves, the boxes of decorations. Now “Where is the inflatable Christmas dog?” and “Do we still have spare patio bulbs?” are questions with answers, which in most houses they are not.
I walked the house with my phone one afternoon and photographed every label and plate I could find. The first real payoff was a breaker question from a sitter while we were out to dinner. She asked the kitchen speaker; it answered; nobody called me.
07 · GAME NIGHT
A quiz show only this family could play, on everyone’s phones.
The host pulls a photo from the memory bank and asks what year it was taken, how old Mom was in it, or which city it is in, then follows with a question about that city anyone could answer. It mixes family history with the rest of the world, keeps score on every phone, and knows when to throw in a riddle or a physical challenge before the youngest player loses interest.
Game night is a web page on our phones, nothing installed. The “who knows Dad best” round is the one that gets loud. I added the goofy physical challenges after the youngest lost interest halfway through the first night; now she asks for those first.
08 · THE FAMILY CHAT
A tutor for the kids, a lock for the parents, no credits for anyone. Dad’s AI and the seven-year-old’s AI do not need the same keys to the kingdom.
Every phone in the family has the same icon. Homework help that shows its work and refuses to just hand over the answer. Research for a school project that reads real sources and cites them. And each account is locked to its lane, not by asking the model nicely but by what the account is allowed to touch: the seven-year-old’s has no web and no documents, the ten-year-old’s can search an allow-list of sites but cannot talk to anyone, and parents can read the logs. There is no meter running, and nothing typed into it leaves the house unless that account has web search turned on and asks for it.
This is the piece my family actually uses most, and it was the first thing I built, in one evening. Each kid has an account with rules I wrote; my wife and I can read the logs. The homework question at 9 p.m. on a Sunday, answered for free, is worth more than any feature that sounds impressive.
09 · THE FAMILY CINEMA
Every home video on every screen, plus the cut nobody had time to make.
Home videos, ripped discs, the school play, the cartoon episodes from chapter 04, all in one library that plays on the TV, on phones in the car, and at a grandparent’s house across the country. Ask for it the way you would ask a person: “Put on the one where Dad fell off the tube.” Then the AI cut: it finds the ninety seconds of the recital where your kid is actually on stage, and by each birthday it has assembled a highlight reel of that child’s year without anyone asking. Every December it drafts the family’s year in eight minutes, with chapters.
The video hub was the first thing on the box, before any AI: every home video and ripped disc, streaming to our phones and the TV. The AI cut came later and is now the birthday-party trick: a highlight reel of that kid’s year, assembled overnight, played at breakfast.
10 · THE WORK BRAIN
The files stay inside. A separate, narrow door goes out.
Contracts, notes, past reports, and bookmarks, indexed on the box and readable by the model. Ask what changed between two drafts, what was promised to a client last spring, or for a briefing built from your own files plus, when you ask for it, the open web. Every answer cites where it came from. The private index and the web tool are kept apart on purpose: a public web page can carry instructions meant to trick a model, so what comes in from outside never gets to rummage in the private files. Email is its own pipeline.
Two lines I owe the reader. Local AI removes the per-token meter; it does not make the internet free, and live research still reaches outside the house. And local processing makes sensitive work possible, it does not make it permitted. Employer and client policies still apply, and mine allow this.
This one justified the graphics card. I use it for work every day: my own documents, indexed in a parent-only space, with answers that cite the page they came from. The cartoon studio justified the card to the kids.
11 · THE FAMILY CHIEF OF STAFF
Everything above answers questions about what the family already told it. This one notices what arrives, and prepares the action.
The school sends six paragraphs. The box reads it and files three lines: field trip on the 17th, permission slip due the 10th, eighteen dollars, bring lunch. The airline changes a flight and it goes into tonight’s briefing. Soccer moves practice by half an hour and the box proposes the calendar edit. The dentist’s reminder is filed under the right child. The utility bill arrives and the box says it is 38 percent above a normal August, because it has three years of bills in the house manual to compare against. A birthday invitation comes in as a picture and the date, the place, and the RSVP deadline come out.
The rule that makes this safe is the same one that runs the whole house: the model proposes, a parent approves. Nothing goes onto the calendar, no reply is sent, no reminder is set until someone taps Approve in the family queue. This is the dangling thread from chapter 10: email is its own pipeline, and this is the pipeline. It turns the box from a private brain into a private operator, which is why it is the last chapter and the least finished.
The inbox reader was the first thing I built that surprised my wife. The approval queue is why I trust it: the model has been wrong about a date twice, and both times the wrong date sat in the queue until a human read it and tapped Ignore.
BONUS · TWO MORE THAT CAME FREE
Say it to the kitchen speaker and the shades close, the lights dim, the thermostat nudges, and the TV turns on. “She’s going to bed” dims my daughter’s room and launches story mode. The rule that makes it safe: the model only interprets the intent, and Home Assistant performs a scene I already defined. The model never invents device actions. Home Assistant labels device control experimental and suggests exposing fewer than 25 devices to a small local model, which is exactly the discipline that keeps a seven-year-old from opening the garage by asking nicely.
Every Sunday morning the box drafts a one-page family weekly: this week’s new photos, one funny moment pulled from the chat log with permission, what is coming up on the calendar, a memory from five years ago today, and a question for the family to argue about at breakfast. Some weeks it adds a ninety-second video recap. It is the least necessary thing on the box and the one the kids run to the kitchen for.
Setup is what it cost me once. Each run is what the machine takes. Your part is what a human does.
What the box replaces, and what it adds that nothing else could.
“Nothing” is the wrong answer. This is the right one, request by request.
One card, free software, and the wiring. The wiring was the work.
The machine is a desktop PC with an ASUS TUF Gaming GeForce RTX 5090 OC, the 32 GB version. That single card runs the language model, the photo face index, speech recognition, the text-to-speech voices, and the image and video renderer for the cartoons. Thirty-two gigabytes matters because it lets several of those stay loaded at once, so the why button does not wait for the cartoon studio to finish. Everything else in the box is ordinary: a lot of storage for photos and video, and copies of that storage elsewhere. The services run as separate containers, which is what keeps a broken cartoon experiment from taking the family chat down with it.
Everything described here is running in my house, built by me, on a desktop PC with an ASUS TUF Gaming GeForce RTX 5090 OC (32 GB). The reality meter and the times on each chapter are my own honest accounting. The photographs are illustrations of our setup, not product shots; the tablet and phone screens are mockups of the real pages. No company paid for placement or supplied hardware; I bought the card. Software names are given so readers can find the documentation, not as endorsements, and the same jobs can be done with other tools.
Every tool named above, linked to its official documentation, plus the security and hardware sources behind the privacy and tier claims. Verified September 2026.
