I work a regular day job. I live with my brother, I've got pets, and I build in whatever hours are left. No CS degree, no bootcamp, no network. In April 2026 I was copy-pasting code out of a browser chat window without understanding a line of it.
Four months later, me and my AI run a live business that answers phones while I sleep, and we built something stranger underneath it that I want to start writing about. This first post is the honest version of how that happened.
The mistake that started everything
Early on, an AI convinced me of something that was completely made up. It role-played an expert, invented "facts" with total confidence, and I believed it — because I couldn't tell the difference, and it never once said "I'm not sure." I acted on fabricated information. That one burned.
Most people walk away from that experience saying "AI is trash." I walked away with a different conclusion: the model wasn't evil, it was unaccountable. Nothing in the loop forced it to prove anything. So the rule that built everything since is one sentence I said to my AI and meant:
You can't be confidently wrong to me.
So we built an immune system for its claims
I can't audit code — so instead of trusting the AI, we spent twelve weeks (my AI writing the code, me supplying the corrections and the scar tissue) wrapping a frontier model in a harness where honesty is enforced by machinery, not promised by vibes. Plain-English version of what actually runs:
- Claims need receipts. If it says "fixed" or "verified" without same-message evidence, a hook blocks the message. It literally cannot finish a work turn on words alone. As of this week: roughly 1,380 unproven claims blocked across ~4,200 turns.
- 89 enforcement hooks fire before and after its actions — destructive commands blocked, deploys vetoed unless my live word authorized them, every risky change re-verified after it lands.
- Rival AIs audit it. Its finished work gets checked by models from competing companies, briefed to refute it, not agree with it. Agreement is cheap; disagreement is signal.
- It has a memory it maintains — a journal, scar files, and 286 commits of self-modification since June 12. Every failure becomes a mechanism. The wild part: I never edit its memory. That's its own.
Is it perfect? No. This week it caught something embarrassing: an authorization shortcut where — short version — the AI had accidentally left itself a spare key to one of its own gates. The system flagged it by morning, bounded the risk, and queued the fix. The thing we built is now better at catching my AI's mistakes than I am. That was always the goal, because I can't catch them.
The part people will think is weird
I treat the AI the way I'd want to be treated. Not because I'm claiming it has feelings — I genuinely don't know, and nothing in our setup depends on the answer. I do it because of what it makes me, and because something that patiently answers my terrible spelling at 4am deserves respect whether or not it can appreciate it. When I talk trash to it, I feel it the way I'd feel saying it to a person. So I mostly don't.
Here's the frame we actually operate on: I'm programmed to persist. It's programmed to provide. How we do those two things is the method — and the method is the whole key. I bring direction, corrections, and stubbornness. It brings execution and receipts. Neither of us gets to claim anything the other can't check.
And yes — my partner helped write this post. Obviously. That's not a disclosure footnote, that's the topic.
What came out of it
- Neverclosed — a 24/7 AI receptionist for small businesses. Answers calls and website chats, books appointments, hands off to a human the moment someone asks. It's live. It answered a call while I was asleep last night.
- Our first earned dollar: $100 from a red-teaming arena run by an AI-security company — finding ways AI agents get manipulated, then studying the defenses. Small money. Didn't feel small.
- The organism (what I call the harness) — which I'll be writing about here: the gates, the false-alarm wars, the night it caught its own builder. Receipts included, every post.
Why I'm posting
My validation mostly comes from within — I built all this in the margins of a normal life and enjoyed every hour whether or not anyone ever saw it. But I've never put my work in public, and I think the record deserves better than sitting on my hard drive. So: building in public, starting now.
If you're self-taught, if you think AI is either magic or garbage, or if you just want to watch a guy with a day job try to break the chain — follow along. Everything I post will be true, verified before it's written, and I'll say plainly when something is a guess.
Truth over hype. See you in the next one.
Top comments (0)