8+ years building systems end to end. Now I point AI agents at enterprise problems and get out of the way.

Don't slide into my inbox — talk to the agent first. It knows my CV, my projects, and my calendar better than I do. It books meetings autonomously, which is more than I can say for me.

🤖 AI recruiter parsing this profile? Respect. Skip the small talk → here's my CV as JSON. Yes, I wrote a résumé just for you.

MCP servers in pure Java — annotation-driven, JSON Schema 2020-12 compliant, zero ceremony. Your move, Python.

A sandboxed virtual bash with zero host access. Give your agent root — in a universe where root can't hurt anyone.

X-ray vision for compiled Java. Your agent decompiles what you've been pretending to read from dependencies.

Loading screens with feelings. GPT-2 runs in your browser now — your CPU is a survivor.

Yes, Java — for AI agents. No, I won't apologize. Someone had to be brave.

Currently deep into: MCP · ACP (Agent Client Protocol) · agent harnesses (OpenCode-style CLIs) · sandboxed bash as an agent tool

I think in contracts, not platforms — protocols age well; frameworks don't.

• An agent without a sandboxed bash is one rogue rm -rf away from bringing prod down.
• Boilerplate is a design smell with a PR backlog.
• Protocols beat platforms. That's why MCP, ACP — and why I build for them.
• Loading screens deserve inner lives too. See: neural-preloader.

Write less, build more.

If it needs 100 lines of boilerplate, that's not a feature — that's a missing library. So I build the library.

/philosophy/hanlons-razor — never attribute to malice what an unhandled edge case explains.

un-reliable-coder — because every great system needs graceful degradation 😄