Human approval gates
Sending, spending and deletion require my recorded approval.
How I shipI use AI-assisted engineering to build infrastructure and tooling: an agent harness, release gates, MCP server vetting and model routing. My AI work uses planted-bug tests, review by a model from another lab and human approval for sending, spending and deletion.
Sending, spending and deletion require my recorded approval.
How I shipOne recent gate passed 35 of 35 tests and caught all 13 bugs we planted before it went live.
How I evaluate releasesMy agents and I built toolwitness.ai: it starts every new MCP server in a no-network sandbox and records what it does. A tool is listed only after scripted checks and independent AI reviews.
Read the toolwitness.ai case studyMy agents and I built an agent workflow harness. It gates sending, spending, deletion and release, and every change gets a review by a model from another lab. In one week (Nov 10 to 16, 2026) it dispatched 1,612 agent jobs; 1,308 ran to completion and 268 were recorded as failed.
| Provider | Tool or model |
|---|---|
| Anthropic | Claude Code |
| OpenAI | Codex |
| Gemini | |
| xAI | Grok |
Python, PHP, TypeScript, JavaScript, SQL and Bash.
I run self-hosted Git, local model inference and backups with sampled restore tests.
LLM evaluation, guardrails, planted-bug tests and model routing.
My job pipeline rejects any answer that cannot quote its source word for word.
My agents and I built toolwitness.ai: it starts every new MCP server in a no-network sandbox and records what it does. A tool is listed only after scripted checks and independent AI reviews.
I run automated newsletter sites Monday, Wednesday and Friday. Each issue passes automated checks before it is sent. Read the case study.
My router picks the lowest-cost capable model for eligible AI helper tasks.
Every safety control I ship gets planted-bug tests and a second-model review.
My job pipeline rejects any answer that cannot quote its source word for word.
Every change gets a review by a model from another lab.
In a GitHub Enterprise Cloud lab, for Copilot Business, I set new features to off by default, turned off web search, outside indexing and bring-your-own-key access, blocked public-code matches and turned on metrics.
Founder and Principal Engineer at IM-Empowered Inc. since 2010.
Power Organics, 2000-2023 (remote from 2010).
B.S. Computer Science, Business Administration minor, Southern Oregon University, 2010.
Senior AI Platform Engineer. Remote preferred; Arizona hybrid case by case.