Standing on the renewal screen, it occurred to me that I could just build the thing myself. I cannot write code, but I can make judgments, so I split those two apart and cleared three problems on the way to a blog that finally fits.
I got all the way to npm chasing a one-line install, and then an independent audit scored it 42.7 out of 100. Here is the two-month report card and the nine things left to fix, in priority order.
I asked for a deck and got back more than I asked for, because it had pulled old meetings and past decisions out of the brain. Records compound. Past the two-month mark, output quality is decided by what has accumulated rather than by the prompt.
One unit has several sizes because each counts a different boundary. Exclusive area first, shared space layered on, applied to the 84㎡ standard and Olympic Family Town.
The system is called Neurain. Split it into three parts, knowledge, operating rules, and a swappable AI host, and the structure snaps into focus. I took apart nineteen comparable products, and here is what I copied, what I refused, and what moat is left.
Ask one question and search runs four stages over three seconds. Split the words and widen them with synonyms, pull candidates from 4,700 documents, filter down to real evidence, then compose the answer on top of it. Answer quality is decided by the evidence, not the model.
I state what I want in Korean, Claude builds it, Codex attacks it, and I keep only the rulings and the dangerous buttons. For someone who cannot read code, trusting the output means measuring behavior rather than reading the source.
Karpathy's LLM Wiki note flips the roles: the wiki is maintained by an AI, not by people. A coding tool that works inside a folder you point it at turned out to be the right vessel. What I changed rather than copied is what became Neurain.
Running an AI as a Telegram assistant widened what I could hand off, but conversations were forgotten, sessions never talked to each other, and switching models reset everything. The bottleneck was not model intelligence. It was the absence of a structure that holds context.
I laid out my own study order: capitalism, money supply, bonds, rates, real estate, stocks, commodities, coins, geography. Each step explains why it sits there.
