I build AI systems that build themselves — and that you can actually trust. Fifteen-plus years architecting enterprise automation, now going deep on agentic AI: multi-agent orchestration, evidence-disciplined workflows, and the governance layer that makes AI survive past month 3.
A two-agent system: a research agent pulls live AI trends and feeds them as opportunity signals into a self-evolving capability engine over a real A2A HTTP mailbox protocol. The engine detects gaps, mutates, validates, and keeps only what survives. It failed eight times before it ran — and it learns from its own failures.
A three-role workflow that makes AI research checkable instead of merely fluent: a router turns a goal into a falsifiable brief, a research agent labels every claim by source and date, and an adversarial skeptic attacks the claims, math, and assumptions — and can demote any of them before a decision is made. The skeptic can weaken a claim but never strengthen one. Automated end-to-end with Claude Code.
Confidence should be earned by provenance, not by how convincing the output sounds. I label what's verified, demote what's stale, and refuse to claim what I haven't measured. Every design choice traces to an observed failure — not theory.