Ian L. Paterson
Practitioner notes on AI agents, cybersecurity, and building public companies in Canada.
Ian’s personal blog.
-
Startup advisory board: build a portfolio, not a committee
In shortI set the pool before I picked the people: about five percent of the company for the whole advisory ecosystem, two points of that to a single anchor. The…
-
How I built a private LLM router: one boring lookup table decides what leaves my network
My LLM router decided privacy by matching keywords in the prompt. It leaked, and I tore it out the next day. What replaced it is one lookup table you can…
-
Best LLM for Coding in 2026: Picks by Use Case
In shortTwo of the 15 models cleared all 38 tasks at 100 percent quality, and one of them costs more than three times what the other does. Sonnet 4.6 runs…
-
Linux OOM Killer vs SSH: How to Stop Getting Locked Out of Your Own Server
When the Linux OOM killer fires, ‘sshd is running’ stops meaning ‘you can log in.’ Four systemd and sysctl fixes that keep your login path alive through a memory storm.
-
I Strapped a $15 Sensor to My GPU. It Read the Load Backwards. Here’s How to Calibrate It.
A $15 Zigbee sensor on an open-air GPU reads load backwards in the fan airflow. An LLM falls for the same trick. Calibrate against nvidia-smi to fix both.
-
My brand-new SLZB-07 Zigbee coordinator shipped too old: fixing “EZSP protocol version 12 is not supported by host”
The SMLIGHT SLZB-07 connected, spoke the right protocol, then quit: EZSP version 12 not supported. Here’s the flash fix and how to pick the right channel.
-
Why docker image prune Won’t Fix a Full Root Disk: The Containerd Snapshotter Trap
docker image prune skips tagged images, and the real hog hides in /var/lib/containerd. How I traced a 99% root disk to the containerd snapshotter and fixed it.
-
Self-Hosting RustDesk on Tailscale Without Opening a Single Public Port
Run rustdesk-server-oss behind Tailscale with zero public ports, plus the macOS root-daemon config fight that ate an hour and nobody has documented.
-
Why Your llama.cpp Benchmarks Are Wrong: GPU Architecture and Real Numbers
In shortThe RTX 2060 SUPER came up clean and generated 52.8 tokens per second, which felt respectable until I noticed the binaries had been compiled for the Maxwell card it…
-
I Built a Honeypot to Catch Prompt Injections in Claude Code (Here’s What It Caught)
I built a honeypot canary that screens web content with a deliberately gullible LLM before my AI agent reads it. Here is what it caught.