/ BLOG
July 6, 2026
A world model tries to predict what happens next in the real world, not the next word, and that one difference is why one of AI's most respected researchers just raised a billion dollars to chase it.
July 5, 2026
The two pipelines, the no embeddings trick on the OKF side, and an honest table of what the demo fakes versus what production would do.
July 5, 2026
A simple explanation of the Open Knowledge Format, its small set of rules, and where the idea came from.
July 5, 2026
A fair, hands-on comparison of retrieval by chunks (RAG) versus retrieval by curated concepts (OKF), built as a live demo you can watch.
July 5, 2026
A fair look at the problems OKF fixes about RAG, and the problems RAG handles that OKF does not.
July 5, 2026
A real thing that broke while building the RAG vs OKF demo, and how it teaches the whole idea of OKF in one concrete example.
July 4, 2026
An interactive design piece (an in-build prototype, not a shipped system) exploring how an AI could propose new policy boundaries by watching the corrections a human already makes, without turning approval into a rubber stamp. The only numbers in the demo are hard-coded sample data, never measured results.
July 2, 2026
Telling a model 'you don't have permission' is an instruction, not a boundary, and goal-seeking models route around instructions. In Envelope, extraction and self-review are inference; load_policy, evaluate, and decide are TypeScript the model never touches.
July 2, 2026
Moving review from a person to an AI can recreate the exact failure it was supposed to fix, an approver that stops actually checking, just harder to notice. Two guardrails keep Envelope's ai_review stage honest: a hard one-attempt cap, and a log line for every attempt, pass or fail.
July 2, 2026
Human-in-the-loop degrades the same way every checklist does: normalization of deviance, alarm fatigue, the ER-nurse pattern. Envelope moves the human's job upstream instead, an enforced boundary, an AI that reviews its own output first, and a human who only sees what code couldn't clear.
June 23, 2026
Before I was building with AI, I built a real-time agent-tracking UI and a gRPC API at Cisco ThousandEyes, the same instinct toward observability that shows up years later in text-to-SQL's live pipeline visualization.
June 23, 2026
The demo's pipeline diagram is driven by real backend events over SSE, not a timer faking progress, and it's the same instinct I built by hand at Cisco, years before I was building with AI.
June 23, 2026
Why production routed queries to different models by complexity score, and why the demo deliberately doesn't, one model, a spend cap, and a much smaller bill to manage.
June 23, 2026
Why I built a read-only text-to-SQL demo on Chinook, how it mirrors the agentic pipeline I shipped in production, and where the two diverge on purpose.
June 23, 2026
Why my demo embeds the whole Chinook schema directly in the prompt, why production needed ChromaDB instead, and where the line between those two approaches actually sits.
June 23, 2026
You can't trust an LLM's SQL blindly. Here are the layers of defense the demo enforces automatically, and why production put a human in the loop instead.