Salesforce architect · Agent systems

I build the parts that make AI agents safe to hand real work.

Memory that holds the right facts, a second agent that tries to break every result, and guards that wait for a human before anything touches production. All of it runs in my own work, every day.

See the workGet in touch

Selected work

Six systems, all live

  1. 007 · VerificationQA Loop With An AdversaryBefore my AI says done, a second agent tries to break it.
  2. 001 · ContextThe Dominio RouterMy AI only reads the parts of my life a prompt touches.
  3. 066 · Live assistA Second Brain In MeetingsDuring client calls, my screen shows cards as people talk.
  4. 047 · MemoryThread notes systemEvery work item gets its own memory file, with a checker.
  5. 016 · OperationsA Board For AI SessionsTwelve AI sessions open at once, all on one board.
  6. 015 · DesktopI Replaced The Windows TaskbarI replaced the Windows taskbar with a shell built for AI sessions.

About

I'm Henry, a Salesforce architect in Medellín. I design and integrate Salesforce systems, and I build the agent tooling around that work.

Most AI demos stop at the answer. The hard part is everything after it: what the model is allowed to see, how you know its result is right, and what stops it from doing something nobody approved.

That is what I work on. I run a dozen AI sessions at once on my own setup, and each project on this page exists because I needed it there first. The code for most of them is public on GitHub.

  • Salesforce architecture and integration
  • Agent systems on Claude Code: hooks, skills, MCP servers
  • Guardrails, verification loops and human approval steps