I architect and build AI-assisted software delivery systems, and I bring that to enterprise and public-sector teams where reliability, security, and governance matter: contact-center platforms handling about 100,000 calls an hour, regulated healthcare and state government work, and now agent pipelines with independent verification. My approach is simple: pin down the problem, write down what done looks like, let AI agents do the heavy lifting inside clear limits, and never let the agent grade its own work.
Recent work includes Ratchetloop, where AI agents write code and independent checks review it; Strategy Assayer, which tests whether a result is real or luck; and Meetup Finder, a live app that finds a fair place for a group to meet.
Before this, I built production systems at Verizon for enterprise and public-sector clients, including Kaiser Permanente, Walmart, and state government. Published research, five granted speech-recognition patents, industry recognition, and AWS ML / AI certifications are documented in Technical Credentials.
Selected Work
Ratchetloop - Auditable AI Software Delivery
In plain terms: AI agents write the code, independent checks and a second AI from a different model family review it, and a person decides what ships.
A local-first software-delivery pipeline for AI coding agents. A bounded task goes to a coder in an isolated Git worktree, deterministic checks validate the change, a different model family reviews the diff, and Ratchetloop leaves a branch and a full record of what happened, so a person can decide whether to merge it.
Enterprise & Public-Sector Delivery
Phone and chat systems for large organizations with strict security and compliance rules. Delivered at Verizon, 2012 to 2024.
- Kaiser Permanente: technical lead on a configurable VoiceXML call-handling framework for roughly 100,000 calls per hour, with CTI screen pops, regional customization, and HIPAA constraints. The initial delivery led to multiyear contract extensions.
- State agencies and healthcare: Dialogflow, NICE CXone, NICE inContact, and Genesys integrations with bot-to-agent handoff, delivered through public-sector security review. A FAQ-bot proof of concept helped retain a $5M annual contract.
- Walmart: IVR for the primary customer-support number on NICE inContact, moved to the new platform on site without interrupting service.
Strategy Assayer - Computational Research & Evaluation Platform
In plain terms: it separates results that are real from results that are luck, so decisions are not made on noise.
The largest system I have built on my own. It runs thousands of experiments in parallel, then puts the best ones through harder tests (held-out data, random controls, and corrections for trying many ideas at once) before any result is trusted. A 679,140-point search finished in under three minutes instead of an estimated 88 hours. The same approach fits any problem where a promising result has to hold up.
Deep technical case study · Public orchestration-engine slice
Meetup Finder - A Fair Place to Meet, No Sign-Up
In plain terms: a group enters where everyone is, and the app suggests a fair place to meet, with no sign-up.
The design starts from three privacy boundaries: no accounts, no app install, and no location tracking. It computes the meeting point itself rather than calling paid map services. Those three rules shaped everything else, and the app went from idea to a live beta.
View the broader project portfolio →
Field Notes
When Code Becomes Abundant
An argument about what happens when code stops being scarce: binaries become machine-readable documentation, proprietary vendors retreat toward the cloud, open source gates cheap contributions, and specification, evidence, validation, data, and judgment become the valuable assets.
A Discipline for Building Production Software with AI Agents
How I keep AI agents building the right thing: decisions written down, facts checked before they're relied on, tight limits on each agent, and review by something other than the agent.
Putting Qwen3.8-27B to Work on a DGX Spark
Experiments on how much real software work a local model on a desktop AI machine can take over from paid cloud models.
Autonomous Improvement Rate
How to tell whether a self-improving system is actually getting better: how fast it runs, how much each attempt teaches it, and how well it holds on to what it learns.
Production Engineering Track Record
- Enterprise and public-sector delivery at Verizon (2012 to 2024): Kaiser Permanente, Walmart, and state government environments.
- Speech and voice systems before Verizon: led delivery of an email-by-phone platform serving 50,000+ daily users, and led VoiceXML conformance testing for the VoiceXML Forum.
- Published research: IEEE fuzzy-logic control paper; NASA conference and international workshop publications on genetic algorithms; U.S. Department of Energy technical report based on Duke M.S. thesis research.
- Patents: co-inventor on five speech-recognition patents across U.S., European, and Chinese grants.
- Industry recognition: principal developer of the Davey Tree voice application, recognized by Speech Technology Magazine as a Most Innovative Solution. Also one of many developers on Verizon's National Operations Voice Portal, which won the same award.
Supporting references are in Technical Credentials.