You have watched this happen three times this year.
Somebody generates an application in an afternoon. It looks right. It has a login screen, a dashboard, a chat box. Then the ordinary questions arrive. Who gets billed. Who can see whose data. What happens when it falls over on a Sunday. Who answers the ticket. The thing stops being software and goes back to being a screenshot.
That is a foal. It cost nothing and it cannot finish a race.
We are a very small shop. We breed the whole horse.
One spec in. The whole horse out.
One specification goes in. Out comes the data model, the business logic and its APIs, the application, users and roles and billing, the infrastructure it runs on, and the operations console you watch it from.
Then the part that costs money for the next five years. It comes with its own support: agents that follow your development lifecycle, working in the tools your people already use. Not a chatbot on the side. The same record the build ran on.
5:35 · narrated by Sue, PractiCAI's coordinator · fourteen real product screens, none mocked up
Why a small shop can do this
The expensive part of software was never the typing, and the typing is the part that collapsed. Consequence did not collapse. Writing a change to shared infrastructure takes forty seconds now. Executing it is as irreversible as it was in 2015.
So the job moved. The question is no longer whether you can build it. The question is whether the thing in front of you is true. And the failure to be afraid of is not the one that goes red.
Three from one week on our own estate. All three looked like success.
- A readiness board read seven of seven green. Three of the seven were false.
- A unit proved its manifest by comparing it against a copy of that same manifest. It passed. It reported no drift because there was nothing to drift against.
- A film shipped in the wrong voice at the wrong sample rate. Nothing on the machine said so. Every render call had succeeded and a file existed at every path.
The symptom also sits nowhere near the cause. A private subnet routes to a small machine that has been switched off since September. What you see is an image that will not pull. What is true is a stopped instance three layers away. The cloud console calls that instance healthy the whole time, because it is.
A person loses a day to that. A team loses a week. No dashboard raises a hand, because every individual system is working correctly.
So what a small shop runs on is one record that everybody is on, people and agents alike. It carries four things: when it was true, what was known, where each thing is, and whose part it is. Lose one and you are holding an anecdote.
An agent needs that record more than a person does. A person fills gaps with judgment. They
have seen this before, they know that box is flaky, they go and check. An agent has no
yesterday and no colleague. It fills gaps with confident nonsense and acts on what it was
told. Writing a record good enough for an agent forces it to be honest, because a human will
forgive status: done and an agent will ship on it.
The four things, drawn. Lose one and it is an anecdote.
What that looks like running
A launch can say no. Seven units, eighty-six checks, one held. The held unit passed
sixteen of seventeen. The screen does not say failed. It names the subnet. It names the
machine that is stopped. It prints the command that fixes it. Six units went. That is a
NO-GO printed honestly instead of a green light printed hopefully.
A Big Bang opens one epic into the work. On our RealWorld Conduit build: ten phases, one hundred and six work units, eight hundred and forty-eight hours modelled, a branch and a pull request per phase. The published RealWorld suite runs thirteen of thirteen, one hundred and fifty-four requests, nine hundred and twelve milliseconds. We do not count the first run as a zero. It was measuring nothing.
Each scar becomes a check that says its own name. The subnet routes to a NAT instance which is stopped. Here is the command. That was a greybeard’s question, travelling by direct message one recipient at a time, leaving when they left. It runs on every launch now, for anyone, whether or not they have heard of a route table. It arrives at the failure instead of in a document somebody has to know to go and read.
The film at the top of this page generated itself from that record. The narration is measured, the picture is rendered to the narration, and the fourteen product screens in it are screenshots of things that run. Two of our products have no screen. They are named and not mocked up.
The stable
Three divisions off one engine.
Enterprise. StellarView is the flagship, the thing everything else is grown on. Faberi runs enterprise AI operations: every signal that needs a person, every agent running, what the models cost today, and a live fabric map where AWS, Anthropic, your own people and its agents are one architecture. TigeRDS takes the database estate. Thirty-one resources across our accounts, scored for recoverability and waste. Twenty-five of them have no automated backup, which is how we found out they do not.
Consumer. Youvocate puts an advocate beside somebody working through their own health. It organises and routes. It never diagnoses. Little Doowop Snapper turns a recording into a film with the whole record underneath it. Fradulous goes after mortgage fraud.
Content and publishing, which is how any of this gets explained, including this page.
Then the verticals: estate law, land acquisition, rooftop solar, aerial spray operations, healthcare intake, pet care, music education. Every one grown the same way.
And the part that is dashed
House doctrine is that solid lines are what works now and dashed lines are what is next. Same rule in public as in private.
- The cloud unit path exists and has never flown. Units still run as containers on a laptop. The launch says exactly that, in those words, on the row.
- Unit boot, the fix log, resume and the monitor are written down and not built.
- The methodology this argument comes from is on disk and not published, which is its own small illustration of the problem.
- Two of our products have no screen worth showing.
We would rather print that list than a cleaner one. The failures earlier in this piece are here because the record kept them. You cannot retrofit that. You either wrote it down as it happened or you are writing an anecdote about it now.
Come and look
The race talk is at practicai.org. The horses are at stellarverse.stellarview.net.
Horses race. StellarView sprints.
Ask the sceptical version. The most useful question is which of the numbers above you cannot check, and we will tell you.