I launched my platform with an orchestra, not a team

For years, building a platform the size of Fawzooz meant one thing: a team. A designer, a front-end developer, a back-end developer, a security specialist, a tester, a content editor, someone for social media, and months of meetings, hand-offs and waiting.

This time it did not. I launched the platform myself, the way I describe in my white paper: vibe orchestration.

From autocomplete to orchestration

In the paper I describe four eras in how we build software with AI. In each, autonomy moves from the human to the system, and the human's role changes:

The fourth era does not mean letting the machine do as it pleases. It means becoming the conductor: you set the melody, you hear every instrument, and you stop the music when it goes off key.

What did the orchestra build?

Fawzooz is not a brochure page. It is a complete platform in Arabic and English:

• a library of publications and books, read on the site, with PDF editions on request;
• interactive tools and indicators;
• a learning programme with certificates that pass through stages of verification;
• the Fawzoozyat magazine;
• an assistant that answers from the site's own content and cites its sources;
• reader accounts that keep their progress;
• a full dashboard, and a live chat that reaches me on my phone;
• a system that prepares social campaigns and publishes them on their dates after I approve.

Work began on 29 September 2026. In eight days the site grew to more than 2,600 pages in two languages, guarded by 193 automated tests that run on every change before anything reaches anyone.

How does the orchestra actually work?

I do not write the code. I write the intent: what I want, for whom, and in what spirit. Then the agents work:

• An agent that builds: it writes the code, the pages and the design.
• An agent that tests: it runs the tests, captures the pages in Arabic and English and on a phone, and makes sure nothing broke.
• An agent that reviews: it looks for security gaps and errors of logic.
• An agent that prepares content: it turns design files into covers and pictures, and books into reading editions.

Everything then comes back to me as a change request that I see and review, and nothing is merged or published without one word from me: “merge”. In those eight days I reviewed and approved 161 change requests this way.

The technology that made it possible

• AI agents in an isolated environment: they write, test and propose, but they hold neither the keys to publish nor any password. The keys are mine alone.
• Hosting at the edge of the network: the site is served from servers close to the reader, wherever they are, with no server for me to run or maintain.
• A database and storage without servers: certificates, accounts and private books are kept with no operations team.
• An automated pipeline: every change is tested automatically, and once it passes and I give the word, it is published on its own in minutes.
• Automation under human oversight: the assistant, campaign preparation and reminders run by themselves, but at the points of decision they wait for my approval.

What did it save?

I estimated what a project of this size would have needed the usual way, role by role:

In this estimate that is a team of ten for about six months: some 49 person-months, or 7,800 hours (at 160 hours a month).

But the real saving is not in hours. The real saving is that the distance between an idea and its making has become very short. Today I say “I want a page for collaboration with three paths”, and I see it ready for review the same day. That changes how you think: you try more, you step back faster, and you build what you had put off for years.

The paper's principles, applied to the platform

All this may sound like a celebration of the machine. It is the opposite. The more autonomous the agents, the more it matters who leads them; an orchestra without a conductor is noise. So I did not only write about the paper's principles; I built the platform on them:

• Validate at runtime: the paper recommends treating AI-generated code as untrusted until its behaviour is validated. So no change passes until every automated test has run on it and its pages have been captured and reviewed.
• Tier the guardrails: cheap, fast checks run on every change, and the costly human review is called only at the points of decision. I do not review every line; I review what deserves a decision.
• One gateway: as every AI request in an organisation passes through one gateway, everything that reaches people passes through one gate: my approval. No merge, no release and no social post without it.
• Extend Zero Trust into reasoning: the agents work in an isolated environment, and each takes only the minimum context its task requires: the principle of Least-Context Access Control. The keys and passwords never reach them.
• An open intake register: every change is kept with its record: what changed, why, and who approved it. And on social media every post says openly that it is part of a campaign designed and run by AI, reviewed and approved by a human.
• Positive error management: when a release stumbled or something came out other than intended, it was not a failure to hide but raw material for learning: fixed in the next change, with a test added so it does not happen again.

This is what the paper calls governed vibe orchestration: give the machine its full capability, and keep the highest authority for yourself.

Who is this for?

For everyone with an idea they put off because they “don’t have a team”. And for every organisation that thinks digital transformation means hiring more people. Today’s tools let one person with clear intent and disciplined governance do what once took a whole team.

The question is no longer: can the machine build? The question is: do you know what you want to build, and how you will govern it?