Design intent
Two requirements drove the layout. A compromise in one plane should not reach the others, and backups have to survive the thing they are backing up. Both are unremarkable requirements that get skipped constantly, because the flat version works fine right up until it does not.
Platform
- Proxmox on physical hosts, running virtual machines and containers, with Docker for application workloads on top.
- Management, production, and backup traffic separated onto distinct segments with routing restricted between them.
- Dedicated backup infrastructure isolated from the production plane, so ransomware reaching a production host does not reach the restore point.
- Secure remote connectivity into the management plane rather than exposed services.
What runs on it
The workflow automation platform, internal applications and their databases, and local inference. Keeping these on infrastructure I control is the reason several of the automations are allowed to touch tenant data at all.
Local inference
Ollama serves models on GPU-enabled hardware inside the internal network. The automation platform calls it for summarization and classification of operational text.
The point is not model quality. It is that ticket notes, call summaries, and configuration text get processed without leaving infrastructure I control, so no third-party model provider sits in the path of client data.
- Used for summarization, classification, and drafting inside existing workflows.
- Sits behind the same segmentation as everything else on the production plane.
- Frontier models are still used where the work genuinely needs them and the data is not sensitive.