11
DECISIONS DEFENDED across the project pages
LLMOPS · CHARLOTTE NC · BY APPOINTMENT
I DESIGN AI
SYSTEMS.
MODELS SHIP.
I PROVE IT.
The professional record. I build LLMOps systems and prove they work — eval harnesses, trace observability, cost routing, prompt maintenance. Twenty years in regulated software; every decision below is defended on its own page.
WHAT I BUILD
ADVISOR PATTERNS EVAL HARNESSES RAG AGENT ORCHESTRATION TRACE OBSERVABILITY COST TELEMETRY
Selected Work
8 SYSTEMS / SHIPPED & DEFENDEDFlagship
Reasonable UX
eval harnessPlaywright + vision
2W/1L — the cheaper config won, proven by eval. Playwright agent that walks a site, scores the UX, and emits a multi-page PDF. Tiered Haiku scout → Sonnet executor → Opus advisor, with a separate eval harness that quantified which variant actually wins — the cheaper one.
Shipped
RAG Brain
RAGVoyage · ChromaDB
<$0.01 per full reindex · 5 defended decisions. A working RAG pipeline over a personal Obsidian vault. Five LLMOps decisions made explicitly and defended, with incremental indexing via content hashing.
Shipped
QA Agent
multi-modelcost telemetry
84% token reduction (~55k → ~8.8k per run). Automated UX quality-audit agent on Playwright + multi-model scoring. Cut running cost 84% by stripping images from conversation history and shipping screenshots at JPEG-40.
Shipped
Agent Panel
multi-agenteval harness
59 expert/reviewer agents + an LLM-as-judge eval. The panel of Claude subagents that runs my dev environment — domain experts paired with adversarial reviewers, plus skills, hooks, and scheduled automation. The system that reviewed this very redesign.
In progress
Reasonable Basis
tax RAGcitation-grounded
Haiku 4.5 beat Sonnet 4.6 AND Opus 4.7 on false-positive refusal — at ~12× lower cost. Semantic tax research grounded in IRS publications — RAG with mandatory citations and graceful refusal. A 7-metric eval, all-pass, on a self-validating corpus.
In progress
rag-lens
RAGspatial UI
A generative-UI 'flipbook' that drops on top of an existing RAG system — spatial visual browse of retrieved chunks instead of a chat list. The engineering-differentiation flagship.
In progress
red-team-bench
securityOWASP LLM
Red-teams LLM-agent prompts and tool schemas against OWASP LLM01 attacks, calibrated against AgentDojo / InjecAgent with Opus-generated hardening recs.
Shipped
Personal OS Dashboard
SvelteKit · Fly.ioLitestream
SQLite mirror via Litestream, on Fly.io behind Cloudflare Access. Private SvelteKit dashboard reading a SQLite mirror of an Obsidian vault. Modules for projects, daily notes, biometrics, and system health.
$0.52–1.94
COST PER UX AUDIT · reasonable-ux
<$0.01
COST PER RAG REINDEX · rag brain