Live model routing
A live router picks the cheapest capable model for eligible AI helper tasks. In one week it routed more than 3.5 million tokens of work to the cheaper Haiku model.
More on model routingPhoenix · Open to roles
AI Automation Engineer
I build AI agent systems that run real business work end to end, with the guardrails that make them safe to trust: planted-bug tests and review by a second model family.
Funnel and offer optimization
Reworked customers' landing pages and offers to raise opt-in rates and sales per click.
About the email businesstoolwitness.ai
A public agent-tool directory. Each new MCP server is started in a no-network Docker lab and published only after deterministic checks and independent review by a second model family.
Read the toolwitness.ai case studymethod
Multi-agent system.
An MCP server connects the agents to the task tracker.
four frontier model families, plus three self-hosted models
| Provider | Tool or model | Hosting |
|---|---|---|
| Anthropic | Claude Code | Frontier |
| OpenAI | OpenAI Codex | Frontier |
| Gemini | Frontier | |
| xAI | Grok | Frontier |
| Open-weight | Three models | Self-hosted |
stack
What I actually run.
Orchestration
Multi-agent systems
Four frontier model families, with an MCP server that connects the agents to the task tracker.
Claude Code
OpenAI Codex
Gemini
Grok
Local inference
Three self-hosted open-weight models
Local inference for three self-hosted models.
Languages
Six languages
Python, PHP, TypeScript, JavaScript, SQL and Bash.
Python
PHP
TypeScript
JavaScript
SQL
Bash
Infrastructure
Self-hosted infrastructure
Self-hosted Git and backups with restore drills.
Linux
Forgejo
Tailscale
Caddy
nginx
Docker
Cloudflare
MariaDB
Model work
Evaluation and guardrails
LLM evaluation, guardrails, planted-bug tests and model routing.
Guardrails
LLM evaluation
Planted-bug tests
Model routing
production
Daily data pipeline; scheduled email.
A data pipeline pulls more than 6,000 job postings a day from public hiring APIs. The LLM's key claims must quote the posting word for word or are rejected automatically.
My system writes persona-specific email copy, with a false-premise checker that flags promises a reader can't verify.
toolwitness.ai is a public agent-tool directory. Each new MCP server is started in a no-network Docker lab and published only after deterministic checks and independent review by a second model family.
My system also runs 56 automated newsletter sites, each emailed Monday, Wednesday and Friday. Read the case study.
model routing
Routing tasks by cost.
My system's live router picks the cheapest capable model for eligible AI helper tasks. In one week it routed more than 3.5 million tokens of work to the cheaper Haiku model.
standards
How I ship.
Checks are proven before they are trusted.
Automated checks are tested by planting a known bug and confirming the check catches it.
Claims must quote the source.
In the data pipeline, the LLM's key claims must quote the posting word for word or are rejected automatically.
Code review
Code is reviewed by a different model family than the one that wrote it.
who
Founder and engineer for 16 years.
B.S. Computer Science, Business Administration minor, Southern Oregon University, 2010. Founder and Principal Engineer at IM-Empowered Inc. since 2010.
For 15 years I ran my own email and traffic business, selling traffic and solo ads to customers and ad networks.
Reworked customers' landing pages and offers to raise opt-in rates and sales per click. Owned paid media, list building, deliverability, offers and back-end resale.
closing
Open to roles
AI Automation Engineer. Remote preferred; Arizona hybrid case by case.