reason=

ADHD-driven, automation-first builder. I architect AI systems. Claude ships them. I verify against reality.

Six things in production. Eight weeks into a deliberate pivot toward AI as primary domain.

what i build

  • advisor patterns
  • eval harnesses
  • RAG
  • agent orchestration
  • trace observability
  • cost telemetry

What's running

Reasonable UX

  • production
  • advisor pattern

AI-powered UX audit tool. A Playwright agent crawls a site and scores pages with Sonnet and Haiku executors; an Opus advisor gets consulted on ambiguous judgment calls.

QA Agent

  • shipped
  • 84% token reduction

Automated UX quality audit agent built on Playwright + multi-model scoring. Cut running cost 84% by stripping images from conversation history and shipping screenshots at JPEG-40.

Personal OS Dashboard

  • production
  • modular · Svelte 5

Private SvelteKit dashboard on Fly.io, reading a SQLite mirror of an Obsidian vault via Litestream. Modules for projects, daily notes, biometrics, and system health.

Vault Search

  • shipped
  • Voyage AI · ChromaDB

Semantic search CLI over a personal Obsidian vault. Markdown-aware H2 chunking, embeddings via Voyage, retrieval from an embedded ChromaDB with SQLite mirror for logs.

System Health

  • shipped
  • self-discovering

Self-maintaining health check that discovers every automation on the machine — launchd jobs, git hooks, scheduled tasks, OAuth tokens, vault integrity — and reports what to run when something breaks.

LLMOps Curriculum

  • tracker shipped
  • writeups in flight

Self-paced curriculum on the operational side of LLM systems — telemetry, evals, cost routing, model gateways. Tracker reads phases, tasks, and progress live from the same SQLite mirror as everything else; the long-form writeups land through summer 2026.

Who I am

Twenty years in tax software, Enrolled Agent since 2011, no CS degree — AI is the first field where execution beats pedigree, and I'm treating it that way.

I architect; Claude ships; I verify. The loop is spec, build, check against reality. Most of what I publish is either the tool I needed or the writeup of why the tool works.

The through-line is automation-first. If something runs twice, it gets automated. The projects on this site are the evidence.