THE JOURNAL
Notes from the workbench.
Engineering memory. Big models, modest machines. Experiments that changed how I think—and the path that brought me here.
The journey
The long way into engineering.
QA, customer support, project management, frontend, full stack. A winding path that still shapes every system I build.
Read storyAI engineering · Practical security
Security boundaries for AI coding agents.
How a practical permissions review became a general guide to agent security: what to restrict, where human approval belongs, and how to test the controls.
Read storyLocal LLMs · Field notes
30 billion parameters. No GPU.
What running Qwen3 on a six-core, 64 GB machine taught me about active parameters, memory bandwidth, and the difference between fitting and running well.
Read storyBuilding Hunch
Code remembers what. Who remembers why?
The idea behind Hunch: keep the decisions, failed approaches, and context that an AI coding session should not have to rediscover.
Read storyEngineering memory · Evidence
Your agent remembered. Did it actually learn?
Building Hunch pushed me to separate four things that are too easily collapsed: recall, claimed use, verification, and actual improvement.
Read storyLocal LLMs · A failed experiment
I made it stream. It got slower.
A 70.5-second local-model pipeline became a 109-second pipeline after adding progressive output. The missing ingredient was actual overlap.
Read storyRetrieval · Measuring the result
Perfect JSON. Worse search.
My local model produced valid structured output hundreds of times. The harder question was whether any of it helped someone find the right thing.
Read storyLocal AI · Building in public
I gave an 8 GB GPU a film crew.
A local director, a video model, and an 8 GB GPU. From an unconvincing action shot to ten seconds of coffee pouring—and what changed when I made the brief concrete.
Read story