// Build journal
Notes, decisions, and code from Neural Bridge and the other things I'm building. Cross-posted to LinkedIn where they fit.
I made my regulatory AI score reproducible, then made it report its own uncertainty, and it still wasn't a real measurement. A score of 0.5523 looks precise. It was four decimals of false precision over a five-band rubric, blended across regulators who were measuring different things. This is how I made the number mean what it says.
I built an AI that scores regulatory submissions, spent a week making the score reproducible, and called it a win. It wasn't. Reproducible and reliable are not the same word, and the gap between them is where a score you can actually trust lives. This is how I closed it: sampling, a stabilized consistency engine, and a harness to check the score against a human.
When an LLM sits in a regulated workflow, a successful prompt injection is not just a security incident. It is potentially a material misrepresentation, an audit trail failure, or a reportable breach. GRC teams are not treating it that way yet.
Most people building a personal AI system start with the dashboard. They sketch a UI. They argue about dark mode. The lesson I learned the hard way: the back of house matters more.
Memory-poisoning prompt injection isn't theoretical. Twenty-four hours after I published the threat model, a friend planted one in the LinkedIn comments. Here's the structural defense I built.
Three days from scaffold to running system. Thirteen specialist agents, a filing-gate compile pipeline, weekly publishing cron, Korean i18n toggle, agents that ship their own PRs with approval. Plus the discipline not to start V3.
If you build, deploy, or use AI in production right now, you're being regulated by people you've never met. Here's the lay of the land.
OWASP didn't pretend AI was a web app. They built a separate project, a separate Top 10, and they're already on the second iteration. Here's the field guide.
I've shipped a sequence of personal AI projects: Neural, Synapse, Argus. The thinking compounded; the substrate didn't. Here's the bridge.
Prompt injection is ephemeral. Memory poisoning persists. Here's how an adversary can plant a permanent lie in your AI's long-term memory, and what to do about it.
A personal AI substrate. Specialized agents share a markdown wiki memory and a chat-based mobile interface.
Building tools mostly for myself. Sometimes I write about it.