Rami HalabiSoftware developer · Stockholm

Architecture

Private AI Datacenter

A reference design for running AI workflows on premises, in 18 linked views from hardware to tenant offboarding.

See the demo

When
Sep 2026
Status
Reference design; not deployed
Source
Public source
Stack
  • Mermaid
  • HTML
  • JavaScript
  • Python

Problem and idea

Organisations want AI to help run their infrastructure, but they cannot send data out or let the AI hold broad credentials.

The idea: keep customer traffic, estate management and AI inference on separate planes, and let the AI act only through a gateway that scopes every credential and records every request.

How it works

  1. BoundariesEstate, networks and service ownership
  2. PlatformVMware with tenant isolation
  3. AI control planeLocal models and durable workflow state
  4. Execution gatewayScoped credentials and approvals
  5. OperationsPatching, backups, incidents and break-glass

Demo

Architecture explorer with a list of views on the left and the estate boundaries diagram on the right.
The explorer: grouped views, design notes, Mermaid source, and SVG export for every diagram.

Open the explorer

What I built

  • 18 editable Mermaid diagrams, each with design notes and source references.
  • A self-contained explorer with search, zoom and pan, fullscreen and exports, and no framework, backend or tracking.
  • A hardware memory worksheet, an automation acceptance catalogue and an offline ZIP.

Stack

  • HTML
  • CSS
  • JavaScript
  • Mermaid
  • SVG
  • Python
  • GitHub Actions
  • GitHub Pages

Quality evidence

  • Structural validation and browser regression tests run in CI before publishing.
  • Rendered SVGs ship with a manifest of their Mermaid source hashes.

Current limits

  • A design only. Nothing is deployed, and the hardware figures are worksheet estimates.