I'm AJ. I run a tree service, and for about two years I've been building things by directing AI coding agents. No CS degree, and I make plenty of mistakes. I try not to believe an AI just because it sounds sure, especially when it's telling me I'm right.
Anatomy of a hosting breach An attack on 17 domains sharing one account. We stopped it, worked out what happened from the logs, and rebuilt 11 production sites so one can't write to another. The write-up includes what we still don't know and what's unfinished.
AxProtocol and Sentinel My first big AI project, with its git history. The idea I still like is that a system shouldn't be the only one vouching for its own output. Sentinel is a separate auditor that recomputes hashes and checks signatures.
opend-ai Everybody's building a coding agent, and this one's mine. It's uncensored, shows you the model's reasoning as it works, and runs commands in a sandbox with no network. Deliberate mode is my experiment in whether small models arguing with each other can get closer to what a much bigger model would say.
claude-h1ve I built this because every new agent session started blank. It keeps instructions and memory in one git repo so each machine starts with the same context. I ran it on desktops and on my phone in Termux.
AI Toolshed Where I keep my AI tools. It has audit and fix skills that make an agent use a product like a customer would and show evidence, plus a guide on writing agent instructions that hold up.
Keno Engine You can't beat keno odds and I don't pretend to. You can change how your wins cluster. It builds 25 coordinated tickets, improves them with hill climbing, and compares them to random picks with a 10,000-draw Monte Carlo simulation.
MediCat Field Guide Repair toolkits assume you already know which tool you need. This offline guide starts from your problem, explains 607 tools in plain language, and tries to stop you from wiping the wrong thing. It's independent, and I've offered it to the MediCat maintainers.
omahatreecare.com and Midwest Roots MCP My business site, with five free homeowner tools for hazards, tree problems, species, DIY or hire a pro, and cost. They say what can't be judged remotely. I'm also packaging them as a ChatGPT/MCP app.
Courses dontlearntocode (early): the AI is the actor and you're the director. It's for people who've committed, with actual syntax on day one. Retiree Business Blueprint (in development): for adults 60 to 75 building a business from what they already know, with a platform designed for seniors.
Smaller stuff bios-pw: on-device BIOS/UEFI recovery candidates archtutor: catches Debian habits in your shell and teaches the Arch way mtk_gui: GUI for mtkclient TahoeGPT: truck assistant PWA for a friend FieldQuote (early): AI estimating for tree contractors ai-futures-lab: research on AI futures strategies
- Picking what deserves the time. I can usually figure out how to do a hard thing. Deciding which ones to do is harder. I'm practicing setting the decision and a stopping point before I start researching.
- Making money from what I build. I've sold for years, but getting my own products in front of people is a different skill.
- Depth. I'm not a career developer. I read and judge code, and I use tests and adversarial review to catch what I miss.
If you've solved any of these, I'd like to hear how. If you're stuck on something I'm good at, maybe we should work together.
AI: Claude Code, Codex, Gemini CLI, Antigravity, Cursor, MCP, CLAUDE.md and AGENTS.md Web: Next.js, TypeScript, Tailwind, Supabase, Vercel, WordPress, PHP Systems: Ubuntu, Debian, Kali, Arch, Windows, macOS, Termux Search: about 10 years of SEO, mostly local, plus GEO