Swapping Haiku for Jev in my AI twin
I replaced the first step of my twin with Jev, a model that picks instead of writing. Refusals got six times faster, and answers more than a second quicker.
Sep 2026Learnings & experiments
Short notes when I learn one thing, longer posts when I break a big one.
Visiting Malian monuments in 3D
Why I started Monument223, the first model that got Djenné right, and what happened when I launched it on September 22.
Sep 2026How I test my AI twin
77 hand-written test questions, scorecards I can compare, what real visitors add, and what I still don't know.
Sep 2026When my AI twin gets it wrong
The bugs that made my twin crash, go quiet or answer the wrong way, and what each one taught me.
Sep 2026Keeping my AI twin out of trouble
The security bugs I found while building my twin, from a script that could run up my API bill to people signing messages with my name.
Sep 2026How my AI twin works (so far)
Build notes on the chatbot that answers questions about my work: why I built it, the stack, and what happens to your message.
Sep 2026Running open-source LLMs on a homelab
What actually fits on 8GB vs. when you give up and rent a GPU.
Apr 2026