Jesse A. Green / notebook

Hardware and architecture

All personal equipment, shown by role, with what is measured and where the sandbox move stands.

1The lab by role

All personal equipment, shown by role.

Role Hardware What runs there
Desktop workstation Ryzen 7 9800X3D, RTX 5090 (32 GB), 64 GB RAM Claude Code, local general and router models, coordinator agent, video rendering
128 GB AI node AMD Ryzen AI Max+ 395 (Strix Halo), about 118 GiB GPU-addressable 120B-class models, long and failover jobs, sandboxed agent
Laptop AI node RTX 2070 Max-Q (8 GB) Small router-class models, sandboxed agent
Lab server (Unraid) 8-core CPU, no GPU n8n, central logging, self-hosted search, file store
Lab firewall PA-440 on PAN-OS 11.1 AI zone designed, not committed
Overlay Tailscale Tailnet-first access; headless nodes managed over SSH
Small cloud VPS Hosted This site
Lab diagram by role: desktop, lab server, two AI nodes and a small VPS on a Tailscale overlay, frontier models reached through command-line legs, and a PA-440 lab firewall; the AI zone around the two nodes is drawn dashed
Figure 1The lab by role. Solid lines are running; dashed lines are designed, not yet built or enforced. Frontier models are reached through command-line legs: Claude Code first, with a second vendor's CLI as failover. full size, 1,600 px

2Measured on the nodes

A 120B open-weight model (gpt-oss:120b) on the 128 GB node reads a 73,000-token prompt at about 200 tokens/s and writes at about 24 tokens/s at that depth, fully on the GPU. In a test job it passed a code task (20 of 20, plus 4 of 4 safety cases) but missed the bar on long-document extraction (recall 0.71 against 0.85) and scored 4 of 10 on a judgement memo.

A 4B router model on the 8 GB laptop GPU writes at about 64 tokens/s. I have no in-lab benchmark for the desktop GPU, so I quote none.

3Where the sandbox move stands

Piece State
Dedicated AI nodes In place
Per-agent identities Per-agent tokens on the hub and the node agents; the node agents run as their own unprivileged user; firewall accounts created, the four read-only AI logins pending; the desktop agents get their own identities when the desktop workloads move to the sandbox
Scoped MCP facades, one token per client In place
AI firewall zone Designed and staged, not committed
Desktop workloads Still on the desktop: local models, the coordinator agent, CLI execution, video rendering