This is where I document ongoing hardware projects — the build, the stumbling blocks, and the measured results. All from real operation, no theory.
AGX Orin — local AI server
64 GB RAM, 456 GB disk, JetPack 7.0. Role: pure inference node (Ollama) behind Tailscale — no Telegram, no key write access. Config master is a separate VPS; the AGX gets an md5-verified read-only copy automatically. Status: productive — article: AGX Orin as a Tailscale AI Server.
Jetson cluster (Turing Pi 2)
Multiple Orin nodes on a Turing Pi 2 backplane, managed from a dev machine running SLURM. Goal: measure distributed inference and model parallelism — not theory, but watts and tokens/s. Status: under construction, article to follow.
Local-AI measurement series (Ollama)
Identical prompts across 5 models: tokens/s, RAM footprint, power draw in watts (jtop logging over days). First measured values from production operation: qwen3.8:27b on AGX — 0.89 s TTFT, 9.8 tok/s. Status: measurement series running, article with complete tables to follow.
Upcoming projects
- Orin Nano vs. AGX on identical models (fair comparison, identical thermals)
- WoL orchestration: waking the cluster via broadcast from the inference node
- Agent mailbox operation: a 24/7 mail endpoint with least-privilege design