Zoltan Fabry

I spent 15 years as a technical writer, including FAA-regulated aircraft maintenance manuals where a vague sentence could ground a fleet. That discipline translates directly to AI. Agents take your spec literally. The clearer you write, the better they build.

I review every piece of AI output as if it has my name on it. Catching confident-sounding wrong answers, spotting edge cases, insisting on functional correctness over semantic correctness.

Breaking complex problems into agent-sized tasks and coordinating the pipeline. My Instructy platform runs 6 specialized agents with a supervisor that catches failures and retries with smarter instructions.

AI agents fail in specific, diagnosable ways: context windows get polluted over long sessions, agents drift from the original spec, they agree with bad data instead of questioning it, they pick the wrong tool, one failure cascades through the whole pipeline, or they produce output that looks right but isn't. I've hit all six and built systems to catch them.

Production AI needs boundaries, not just prompts. I build content gates, run pen tests on my own infrastructure, and keep secrets out of AI conversations entirely.

Context architecture is one of the hardest problems in agentic engineering. I built Open Brain as a custom MCP server - a persistent memory system with semantic and hybrid vector search that agents access through Model Context Protocol. Shared context, specification enforcement, and session management across multiple AI agents and devices.

Every agent in a pipeline doesn't need the same model. I use smaller, faster models like Haiku for validation and routing, mid-tier models like Sonnet for content generation, and reserve Opus for complex reasoning. Picking the right model per task keeps quality high and costs predictable.