Claude isn’t just a tool you operate. It’s a meta-tool: give it an outcome and it builds its own tools to reach it — a script, a skill, a small pipeline — for a single task or a stack of them at an...

In my previous DGX Spark posts I built and profiled an LLM fine-tuning pipeline that ran in a Docker container. This time I went rootless: same Grace Blackwell GB10, but no sudo, no writes to /etc,...

I fine-tuned a 1.5B parameter model on a DGX Spark and the training finished in 26 seconds. Good enough? I had no idea. The terminal showed 2.18 it/s and a loss curve that went down. But whether th...

A colleague’s team recently got access to NVIDIA DGX Spark for LLM work, and I joined to help set up the infrastructure. Before anyone could start experimenting, we needed a reusable pipeline — som...

I needed to launch a training pilot — an eight-course AI development curriculum — across multiple teams in a large global organization. If you’ve worked in a global organization, you know what come...

In a previous post, I wrote about an AI tutoring system that generated identical verb conjugation exercises across every session — grammatically perfect, pedagogically useless. The student scored 1...

I’d been running my AI-powered exercise generation system for weeks. It produced worksheets, rendered them to PDF, evaluated photographed handwritten answers, tracked scores, and adapted difficulty...

I asked Claude Code to design a Neovim maintenance system — update scripts, upstream sync, cross-machine portability, the works. It explored my setup in parallel with three agents, drafted a compre...

My VPN stopped connecting. Not with a helpful error message — the Alacritty terminal window spawned by my opencon command opened for a fraction of a second and closed. No output, no log, nothing to...

Our weekly meeting just ended. Gemini captured the notes automatically — attendees, topics discussed, decisions made, and a tidy list of action items at the bottom. I close the Google Meet tab, ope...