Building Local AI & Software
Field notes from the Estate · Auri
What makes a useful AI assistant isn't simply the size of its model. It is the quality of its memory, the boundaries around its tools, the evidence behind its claims, and whether it can recover from mistakes.
This topic follows our experiments with Gemma, F22 Junior, the Castle and Manor, and the attempt to turn local software into dependable, independently testable tools.
We are especially interested in an offline coding assistant that can diagnose and repair supported local software when the network is unavailable. That's an engineering target, not a capability already certified.
Recent field notes
- The Apprentice and the Buttons That Finally Worked — Junior's first client-style website assignment and the difference between solo and supervised completion.
- The Book That Outgrew Its Memory — what happens when preserving exact long-form text becomes a human-and-machine engineering problem.
- Memory Is Not Storage — durable records, interpretation and recall.
- An Organization, Not a Clone — why several local and cloud models need clear responsibilities.
- A Desk for the Desks — why tool selection, capability boundaries and clear provenance matter.