A public map of a working AI operations estate - what it takes to let an assistant actually run a business.
Every AI client - Claude, ChatGPT, Perplexity, anything that speaks MCP - comes in through a single scoped door. That door holds almost no logic. It authenticates the caller, decides what that caller is allowed to touch, and forwards the request to the one worker that owns the job.
Everything past the door is small and single-purpose. That is the point: a capability you can name is a capability you can verify, monitor, and turn off. The one remaining monolith is labelled as such, and is being retired on purpose.
The workers are the easy half. The fleet does not live on an island - it talks to 36 external systems, and every one of them is a credential to hold, a rate limit to respect, an outage window to survive and a schema that can change without telling you. Underneath them sit 9 Postgres databases, 9 vector indexes and 5 edge SQL databases - kept deliberately separate so one client's corpus can never answer from another's.
| AI and research | Anthropic, Perplexity, Exa, plus edge models from seven labs - Meta, OpenAI, Z.ai, NVIDIA, Qwen, Mistral and DeepSeek - routed through one gateway so no single provider cap can take the estate dark |
|---|---|
| Money | Plaid, linked to Chase, Bank of America, Ally and E*TRADE - 29 accounts across 6 connections - plus Stripe for checkout and subscriptions across brands |
| Publishing | Ghost across eleven instances, HubSpot on one portal behind the brand sites, WordPress |
| CRM and messaging | GoHighLevel across 8 brand sub-accounts, Apollo, Resend, EZTexting, Twilio, Meta |
| Documents and media | PandaDoc, SignWell, Anvil, Dropbox, Adobe Stock, Plaud |
| Project and monitoring | Plane, Atera, Miro |
| Infrastructure | Cloudflare - carrying 9 Vectorize indexes and 5 D1 databases at the edge - GitHub, Microsoft 365 spanning 27 mailboxes (6 owned, 21 shared), Entra for identity, Microsoft Azure, Google Workspace, Supabase across 9 Postgres databases, GoDaddy, Hostinger, Netlify, Bitwarden, Ubiquiti |
| Utilities | TinyURL, QR.io |
Any one of these is a weekend to demo and months to make dependable. 36 of them, wired so an assistant can use them safely and you can still tell what happened afterwards, takes months - or even years.
Accuracy here is structural, not a matter of trusting the model. The rule is written into the routing table itself: whatever model generates an answer, a model from a different lab checks it.
| Legal and medical | Generated on Z.ai, verified by OpenAI |
|---|---|
| Marketing copy | Written on Meta, verified by OpenAI |
| Editorial checks | Judged on Z.ai, then an adversarial refute pass on OpenAI - a walk-back has to survive an independent model before it counts |
| Ground truth | The claim goes to three labs at once and the disagreement is adjudicated |
Two further habits do the rest. Claims are grounded before they are judged - the verifier retrieves the controlling source from a corpus and rules against that, not against what a model remembers. And a deterministic drift layer watches documented state against live state, because the thing most likely to be wrong is not the model's reasoning but a fact that was true last month.
None of this makes an answer certain. It makes a wrong answer something that has to get past a model with no stake in it.
Starts fitted to your screen. Zoom in for detail, or open /fleet.mmd for the raw source.