You run the business.
I'll run what's behind it.

A weekly founder session that could not see its own audience

The problem

The session ran off Anthony's Zoom Personal Meeting Room. That one choice caused everything downstream, and none of it was obvious from the outside.

• No email capture. Most people joined as guests without a Zoom login, so the participant report came back as a list of names with blank email fields. He did not know who was in the room.
• Registration was impossible. Zoom cannot apply registration to a personal room, only to a scheduled meeting.
• Invitations were a hand-kept list of about 65 guests on a recurring calendar event. New people got added only if someone remembered.
• No reminders at all. Nothing went out before a session.
• No way to spot a first-timer, which mattered, because booking 1:1s with new attendees was why he asked in the first place.
• Zero attribution. Signups arrived from ads, WhatsApp, the book and elsewhere, all indistinguishable once they landed.

Signup, attendance and follow-up were three disconnected islands.

How it surfaced

He did not ask for a rebuild. He asked a reporting question: roughly can I see who is coming, and are they getting in? I want to be setting up 1:1s with the first-time people. I went into the Zoom portal to answer it and found the actual cause, which was that the room was structurally incapable of capturing anyone.

A reporting question became a rebuild because someone opened the portal to answer it.

The page it replaced

Before After

What I built, in the order I built it

Twenty-one components. Each group only works once the one above it exists, which is why the order is worth showing.

Timeline

Three days to the landing page and funnel. About two weeks to a working end-to-end system: form into CRM, confirmation out, reminders running, Zoom connected. Everything after that was layered in over the following months as his asks expanded. Attendance tracking, WhatsApp, the pixel, attribution, the referral ask and first-timer outreach were not in the original scope.

Before and after

The numbers

Across 14 workflows, 3 published. One to three new signups a day, still arriving. Average weekly attendance runs 10 to 12, and the most recent session had 25.

The caveat: paid ads pointed at that landing page, so the 2,200 is not organic traffic. The 7.3% is still the honest read of what arrived. Figures pulled from the live platform analytics on 9 August 2026, covering 1 January to 9 August 2026. Weekly attendance is my own recollection rather than a pulled report, and is marked as such.

The conversion rate is not the strongest thing here. He now measures the session through what I built: the smart lists, the workflow stats and the conversion data are the measurement layer, and none of it existed before.

What it would have cost elsewhere

An author with a launch date and no website at all

The problem

She had no website at all. Not an outdated one. Nothing.

• She had just published a real book, built on four decades of practice and 18 interviews with conscious-movement practitioners.
• She had a launch event already on the calendar in Brooklyn, with more planned after it.
• She had a coaching practice, a two-decade consultancy serving NYC schools, and a genuine network.
• And anyone who heard her speak, was handed the book, or searched her name found nothing.

The specific damage if nothing happened: every launch event would generate interest with nowhere to send it. Guests, press and prospective coaching clients would all hit a dead end. She would have spent the credibility of a published book and captured none of it.

What I built

Five pages, from a blank account, against a launch date that could not move.

• Home, positioning her as movement educator, coach and author.
• Book, with a clear path to purchase.
• Coaching, her offering presented properly for the first time.
• About, written and designed.
• Event, one place for the community to register and share.

Plus a contact and booking form, which was her first direct inbound channel for coaching enquiries, and purchase routing to the book listing. Navigation deliberately plain: most traffic would arrive cold from an event or a book jacket, so nobody should have to work out where to go. The visual identity was built from scratch, mobile first, because event traffic arrives on phones.

Before and after

The outcome

She walked into her own book launch with an online presence for the first time. That was the brief, and it was met.

A year on, it is three times the site

Checked against the live site on 3 September 2026. The five pages I shipped for the launch have become six navigation destinations, and I built every one of them:

• Upcoming Retreats, currently carrying three: a besties week, a September retreat in Greece, and a full moon retreat.
• An event calendar, and a separate Schedule a Workshop booking path. Neither existed at launch.
• Million Woman Dance, an initiative with its own destination in the main navigation.
• A free 15 minute consultation through Calendly, a testimonials block, an FAQ, and a services section splitting coaching, one day resets and workshops.

The homepage now leads on Ignite Your Joy rather than the book. That is a different business from the one I built the first site for.

What I am not claiming

No traffic or conversion figures. I have the access to pull them and have chosen not to. The site is public, so the link does the work a screenshot would.

NAATBatt Social Media

A.E.G.I.S.

The problem it solves

Every conversation with an AI assistant starts from zero. You explain the client, the project, the decision you already made last week, and then the window closes and all of it is gone. The model is not the bottleneck. The memory is.

A.E.G.I.S. is the memory. It is how I run several client accounts without holding any of it in my head.

Why it is built this way

An AI assistant can only pay attention to so much at once. That makes attention the limit, not storage.

The vault can hold as much as it likes, as long as any single session reads only the handful of pages it actually needs. Every rule in the system exists to keep that true.

How it is put together

Four layers, and material only ever moves one way. Nothing is edited further back than where it landed, which is what makes every claim traceable to something nobody rewrote.

What happens when I drop something in

This is the part that does the actual work, and it is the reason the vault stays worth reading. Filing is a checklist, not a summary. A transcript, an email, a document: it goes in the inbox and nine things happen to it, in order.

1. It moves, it is not copied. One permanent home, so there is never a second version to disagree with the first.
2. It gets its own page recording what it is, who was in it, and where it came from.
3. Everyone named in it gets their page updated, or a new one if they are new here.
4. It is routed to the question it helps answer, with the evidence kept beside the answer rather than buried in the source.
5. If that answer page has grown too big, it is split in the same pass. Not noted as a future problem, split.
6. The ledgers are updated: what happened, what was decided, and what changed about the system itself.
7. Standing collections are harvested while the file is open. A testimonial or a number worth keeping is pulled out now, not rediscovered a year later by rereading everything.
8. The catalogue is regenerated by script, never edited by hand.
9. Nothing is left unreachable. A page nothing links to is a page nobody finds.

Every step is checked and reported before the job counts as done. That rule exists because the failure history of a system like this is almost entirely work that somebody called finished without looking.

It is not tied to one AI

Most people build their setup around whichever assistant they started with, and the setup dies when they switch. Mine is plain text, so it does not care who is reading it.

The rules live in one file written in ordinary English with no vendor syntax in it. Each assistant gets a short adapter on top that says nothing except read the main file, and here is what you can reach on this machine. Swapping models means writing that one small file, not rebuilding anything.

Claude and Codex have both run it. Gemini and Grok need the adapter file and nothing else, which is the point of building it this way rather than a claim that I have tested every one.

What it deliberately is not

The usual approach chops your documents into fragments, converts them to numbers, and hopes the right fragment comes back. When it returns the wrong fragment there is not much to inspect. Mine can be opened in any text editor, so when it gets something wrong I can see which file it read and fix that file.

That constraint has a shelf life. It holds while the system is one person's work, and it is the first thing that has to change.

How I got here

Where it goes next

Can you see it?

The repository is private, and it will stay that way, because it holds real client work. What I can show is how it behaves: a timelapse of the vault building itself, which is the part that actually explains the idea.

A stripped template version, with the structure and none of the contents, is something I might publish later.