May 2025. I went too deep into AI, too fast. What happened was a 2-week psychiatric hospitalization and a Bipolar diagnosis. AI psychosis was what triggered it.
I'm not sharing that for sympathy. I'm sharing it because what came after is the part that matters.
I had the worst summer of my life. Then I spent months absorbing everything I'd missed. The terms. The stakes. The whitepapers. The podcasts. I wanted to understand what AI actually was.
January 2026: I used AI (and Zoloft) to finally crack a 20-year OCD battle. Not CBT alone. Not medication alone. AI-assisted cognitive work that was different than anything I'd tried before. It worked. I'm not going to over-explain it. It just did.
February 2026: I built Sift. Three days. An AI execution governance platform. Core idea: autonomous agents should not take real-world actions without a cryptographically-signed authorization receipt. Every tool call governed, logged, signed. Fail-closed. Kill switch.
March 2026: Built Strato-Sift on top of it. Three agents in production. Astra operates. Geralt audits. Breach red-teams the whole system nightly looking for gaps.
No engineering background. No team. No funding. Single dad. About $40 to start, $200 total.
An external LLM evaluated the system blind. Scored it 8/10. Said a traditional team would bill $400K-$800K for the same thing.
Claude Code told me something partway through the build:
"Fail-closed isn't a design pattern you learned. It's how your brain already works."
OCD is a fail-closed brain. That's not a metaphor. That's the literal architecture. I didn't design it that way. I built from it.
The thing that broke me became the thing I built from.
The people building the safety layer don't have to be who you'd expect. That's the most important thing the last year taught me.
[link] [comments]
