I turn vague problems into software that ships and stays reliable.

One command turns a niche plus a city into a qualified, researched, ready-to-send outreach list. The LLM is used only where judgment is needed; every other step is gated, logged, and reproducible. Built to prove a thesis: most repeatable "Virtual Assistant" work is a fixed pattern that a deterministic pipeline runs faster and more reliably than a human.

• Tool-boundary design - each capability is a discrete, testable function (scrape, analyze, generate, research), the exact discipline an MCP server needs.
• Production hygiene - secrets via env vars (never hardcoded), a demo mode that runs offline on mock data, idempotent runs, structured CSV plus markdown output.
• Determinism over magic - the thing customers actually want is reliability, not creativity. Ship the wedge: one command, one outcome, no scope creep.

A personal operating system for getting real work done with AI agents: modular skills, role-based subagents, self-healing error recovery, and guardrail hooks on every tool call. Built to answer one question - how far does a single operator scale by orchestrating agents instead of doing every step by hand?

• Systems thinking - decomposition, separation of concerns, observability, failure handling. This is solution-architecture work, not prompting.
• Reliability by design - errors stack (90% per step is 59% over five). So on any failure: classify, route to a fix (install / permission / credential / retry), re-run before escalating.
• Compounding velocity - reusable skills mean each task starts from accumulated capability, not zero. That is the FDE edge: ship fast because the tooling compounds.