// Build journal
Notes, decisions, and code from Neural Bridge and the other things I'm building. Cross-posted to LinkedIn where they fit.
How a full architecture audit, a User vs. Admin UX split, Gemini 4 Argon role routing, procedural 3D Blender emblems, and a three-phase product roadmap turned Bellwether into a closed-loop regulatory engagement system across DORA, UK CTP, NIS2, and the Korea Network Act.
Why a cute 3D lunar habitat and a heavily governed agent workbench both failed my daily-driver test, and how I rebuilt Moonbase into a first-person Mac mini command deck for GitHub Copilot and Claude Desktop.
Technical architecture of the Moonbase Upstairs Mac mini AI Command Deck, live Copilot and Claude session tailer, first-person dinosaur camera physics, procedural Blender desk assets, and virtual rotating photo frame.
Two prompt changes moved my compliance assistant's refusals by about three points. Swapping Gemma 3 4B for Gemma 4 E4B moved them by thirty-six. What changed between the models, what it looks like in the answers, and the rule the first run failed.
How a small local model answers compliance questions from a governed library of primary regulatory texts: the data flows, the corpus rules, retrieval, generation, the evaluation and judging stack, what it is for, and what comes next.
A Neural Bridge conversation-capture path was dead for ten weeks while every log line said it was fine. Why AI governance frameworks and audits are not built to catch that, and how the canary I built afterwards missed the next one.
Giving Moonbase durable specialist profiles, owner-directed background work, curated knowledge and reviewable memory, without inventing activity or authority.
The implemented contracts for versioned specialists, owner-directed native runs, curated knowledge, reviewed outcomes, scoped memory and history-safe maintenance.
The first end-to-end baseline of my compliance assistant: what a library of primary texts is worth against the model's memory, how my grader overstated failure until I read the answers, and why a prompt change died in 22 minutes.
Neural Bridge's auto-reload watcher exited 1 on every run from May to September without writing a log line, and merged pull requests sat undeployed. The cause was one line of shell in a cleanup function.
Teaching AI security means attacking a real model in front of the room, so I built the room: slides that run the attack, a measured success rate instead of one lucky run, and a recorded clip behind every demo.
Turning an agent office into an original Blender-built lunar workspace, without confusing a cute character with a running agent.
Building tools mostly for myself. Sometimes I write about it.