I'm 8. My toy company runs on AI. Here is the seven-year plan.

WRITTEN TOGETHER BY JOSEPH AND DAD · TECHNICAL SECTIONS BY DAD

THE LONG VERSION · FOR GROWN-UPS AND BIG KIDS

I'm 8. I'm building an AI-powered toy lab. What would you make? Not as a slogan: it is the first thing the website asks you, and it is the whole business in one line. Texas Toy Lab sells 3D-printed toys at markets around Austin (online toys are coming), and the newest thing we sell is a toy that did not exist until you described it.

The Imaginator: imagine your toy. Build it.

The Imaginator is a page where you type a sentence and get back a printable toy. Behind it, an AI draws a picture of your idea, another turns the picture into a 3D shape, and the Lab's own software checks that the shape is solid enough to print, sizes it, estimates the print time, and lets you pick colours. If you want it made, Dad gets the request and taps yes, and it goes onto the print farm. Rather than describe it, here is one from this morning, exactly as it happened.

One real example: a tiny friendly robot holding a star

The idea. "A tiny friendly robot holding a star." Seven words, typed at 6:50 in the morning before school.

What appeared on screen. Forty-eight seconds later, the first picture: a white robot with yellow ear pods, a star on its head and a star on its chest. A good robot. Not holding anything. Under it, the Lab's own report on the 3D shape it had made from that picture: watertight, 56.5 by 80 by 40 millimetres, about 156 minutes to print, one colour.

What needed changing. Two things. The star was on the robot, not in its hand, because the picture-drawing AI read "holding a star" the way a badge holds a star. And the whole toy had come out in one colour, a warm grey, because the shape-maker merged the three parts it found into one piece. I tried the colour step first and turned the whole robot yellow, which the Lab recorded as revision two of the same toy. Then I rewrote the sentence: "a tiny friendly robot holding a big star up high in one hand, star and robot as one solid piece." Ninety seconds later, try two: a yellow robot with a teal star held up in its right hand. Watertight, 65.5 by 80 by 36.6 millimetres, 131 minutes to print, still one colour, which is fine for a first print.

{
  "prompt": "a tiny friendly robot holding a big star up high in one hand, star and robot as one solid piece",
  "status": "done",
  "watertight": true,          # no holes: a printer can slice it
  "size_mm": [65.5, 80, 36.6],
  "print_estimate_min": 131,
  "slots_merged": 3,            # three parts became one piece, so one colour
  "revision_colors": ["#CD9924"],
  "state": "preview",
  "approved_revision": null     # nothing prints until an adult approves a revision
}

The print. At 6:58 I pressed "Print it". That does not start a printer. It puts the toy in the queue and sends Dad the request; a toy only prints after a grown-up approves the exact revision. Dad's tap is pending as I write this, the printer will take a little over two hours, and the finished robot, photographed, is in note two on September 24. If it comes out wrong, that goes in the note too.

Joseph and the AI

Most of the Lab's daily work is done by a crew of AI agents on a rack of computers in our house: one for the website, one for social posts, one for markets, one for the print farm, one for legal, one that keeps the money honest, a boss agent that reads the plan every morning, and an outside second opinion, a Grok Bot from xAI, whose only job is to poke holes in the plan. They write, schedule, watch the printers, answer texts, and file everything they do in a log that cannot be quietly edited.

Who approves what. There are two kinds of permission, and the difference matters. Anything new, anything with money in it, anything that makes a promise to a customer, a partner, the press or a school gets an individual yes: a card on Dad's phone, tapped one at a time. A small list of routine actions runs under rules Dad approved once, in writing, with a short window in which he can still pull any single item: a social post from the queue he already approved, a text reply that only repeats facts from this website, a market-day reminder to someone who texted MARKET to ask for it. Every routine item is still recorded, and the list of rules is reviewed every Sunday. Nothing outside that list goes out on its own. The crew does the digging and the drafting. I still do the deciding, and Dad does the money and the law.

How the Lab is wired up

The print farm is nine Bambu Lab printers on a steel shelf: six P2S machines for everyday volume with four colours each, an H2S for big and shiny prints, an H2C that swaps its own hotends, and the A1 mini that started it all. They live on their own Wi-Fi and their own slice of the network, report every layer to Norman, and a finished print is not a finished job until a human has taken the plate off and counted the pieces. Norman is a rack in a closet: a used Dell PowerEdge control node where the crew lives, and four NVIDIA DGX Spark computers, each the size of a hardback, each with a Grace Blackwell chip and 128 GB of memory shared between CPU and GPU (NVIDIA quotes up to one petaflop of AI maths each). Between them they run a 120-billion-parameter open model, a fast 30-billion one, a vision model, the Imaginator's picture and 3D models, and Buddy's ears and voice.

What leaves the house, exactly. Claims about privacy should be specific, so here is the list. Stays on the rack: everything a child says to Buddy, every Imaginator prompt and model, every draft the crew writes, the printer feed. Leaves the house, because it has to: orders and payments go through Shopify; texts to and from (512) 399-2625 go through Quo, our phone provider; email goes through Resend; the website itself runs on Vercel; and a request from the public internet reaches the rack only through our own small relay, which the rack dials out to. Leaves the house by choice: the Grok Bot review, an outside AI service that reads our plan documents and insurance quotes, never a customer's details and never anything a child said. There is also a cloud-voice option for Buddy that we tested, measured as slower than our own rack, and left switched off. If any of this changes, the change goes in a note.

The plan, as evidence

None of the above is the point on its own. The point is a business that a kid can run and grow, and we wanted a plan we could be wrong about, not a wish. There is a spreadsheet with seven years in it and a file the crew reads every morning. The numbers below are copied from those two files.

The assumptions, so you can argue with them. "EBITDA" here is profit before tax and interest, after the running costs of the business including a vehicle allowance; "margin" is that profit as a share of sales. "Business value" is EBITDA times a multiple, and the multiple is illustrative: small owner-run businesses often sell for less, so the model also shows a version that counts only half of it.

So: in the downside case the plan misses, in the base case it needs the business to be worth something, and only in the upside case do the savings alone get there. Year three margin and the multiple are the two assumptions most likely to break, and the workbook says so on its first tab.

north_star:
  statement: >
    Joseph Jr. reaches $1,000,000 in bankable and business assets by age 15
    (fiscal Y7, 2032/33) through a real toy business he runs, printed at home,
    honest about what is his design and what is licensed, safe for children,
    with his parents handling money and law.
  proof_points:
    - Y3 is the hinge: revenue 250,800 at 53% EBITDA margin makes or breaks the model

year_1:                      # Sep 2026 to Aug 2027
  revenue_total: 38000
  non_negotiables:
    - every customer-facing claim is true (kid-run, printed by a kid; licensed designs credited)
    - COPPA: no marketing to children, grown-up consent at every capture point
    - no new recurring spend without Joseph's approval

Why we are doing this

The thirteen notes

One every second Thursday at seven in the morning (this first one is a Friday, because we could not wait), each with a short version for kids on top and the real thing underneath: numbers, sketches, a real photo and the actual code.

1. 01I'm 8. My toy company runs on AI. Here is the seven-year plan: this noteSep 11
2. 02How the Imaginator turns a sentence into a toy, with the robot from this note, printedSep 24
3. 03The second opinion: how we use a Grok Bot to argue with our own planOct 8
4. 04Buddy gets a body: setting up a Reachy Mini robotOct 22
5. 05Setting up a print farm of Bambu Lab printersNov 5
6. 06The software we built to run the farmNov 19
7. 07Norman, the computer that runs the LabDec 3
8. 08Four little supercomputers: the DGX SparksDec 17
9. holiday break
10. 09The harness: how a crew of AI agents runs a toy business without running awayJan 7
11. 10The Lab's phone line is a robot with a managerJan 21
12. 11The control room in my pocket: the Lab consoleFeb 4
13. 12Ears and a voice: speech in the LabFeb 18
14. 13What we watch, and the six-month review against the planMar 4

Note thirteen is a six-month review, not the year-one scorecard; year one ends in August 2027, and that scorecard comes in September 2027. If you only read one, read that one.

Howdy. See you at the next market.
Joseph ★ (and Dad)