writing

Build logs.

Notes from building local AI systems. The failures stay in — they are the useful part.

By Brian Niceley

001 · July 2026

The evaluator that caught my own bug

My tests were green. An independent adversarial pass found a real fail-closed bypass anyway — and then had to prove the fix.

002 · July 2026

Start with the work, not a GPU shopping list. A sizing guide for matching models, memory, concurrency, and an honest budget.

003 · July 2026

Pull the cable and run the real workflow. Here is what the test proves, what it catches, and what it cannot promise.

004 · July 2026

The useful workflows, the custody advantage, and the jobs where a local model is still the wrong answer.

No publishing schedule promised. Entries appear when the work produces something true worth writing down.