Kyle Bartlett · senior AI engineer · Houston, Texas
15 agents and 52 scheduled jobs ship code to production, roll it back on command, review each other's work and pass jobs to another vendor's bots, with nobody relaying in the middle. I'm the one engineer behind the fleet. Before it, I built an agent platform inside a global consumer electronics company that ran live with five of its national retail customers, and it was still running after I left.
Open to senior AI engineer and forward deployed engineer roles, remote from Houston. US citizen, no sponsorship needed. Recruiters, start here.
[email protected] · code: five public repos on GitHub (opens in a new tab)
Codex reviews what Claude writes before it lands, on every project, every day. Prompt changes go through Weigh, a paired eval with a noise floor and a gate that can say no. Its first run rejected a change that saved 10% of prompt tokens because recall fell inside the noise, and no judge counts until it's been checked against human labels.
Tower sweeps the fleet every five minutes with no model call, opens one incident per failure, asks a single healer for a restart and escalates on a clock. When the fleet's alert texts went quiet for 16.5 hours, I root-caused it across two systems from log timing alone, shipped the fix behind the test, typecheck and chaos gate, and verified it against the next real event. Undo puts the previous version back in 14 seconds with a revert commit so it can't come back.
I've worked in operations since 2014, so I know what a team will adopt and what it will route around. At a global consumer electronics company, 65 people used the automations I built and 45 of them ran them every day; once the first retail customer was live on the forecast platform, each new one took about a week of re-testing rather than weeks. As its US AI lead I set the standards, guardrails and approved workflows, and ranked first of 200 employees in company-wide AI adoption.
All 47 projects · 25 live · 0 beta · 22 lab
The console my fleet runs on: 52 scheduled jobs, one owner agent, Undo on everything.
proofTyped on the live page, a request shipped to production in 7 minutes and rolled back in 14 seconds.
Insurance triage in five Claude nodes; anything under 70% confidence goes to a person.
proofThe 26-case golden set went from 91.2% field and 88.5% decision accuracy to 100 and 100.
Open-source before-and-after eval harness with a noise floor and a gate that can say no.
proofFirst run rejected a token-saving filter: 10% fewer prompt tokens, but recall fell from 24 of 27 to 21 of 27.
Both iOS apps went through App Review end to end, built solo with the agent stack behind this site. Open a listing, install one, and judge the work the way a user would.
I build agent systems, run them in production and answer for what they do. Email is the fastest way to reach me, and a live walkthrough of any system on this page is an easy yes.
Need something like this built inside your own company? That's Bartlett Labs.