Profile
Jeffrey Michael Johnson
AI Engineer // Agent Orchestration // Phoenix, AZ
AI Engineer. Designs in Figma, ships to production, and builds the agent systems that hold the pipeline together.

01 // Who
The short version.
Jeffrey Michael Johnson is an AI engineer in Phoenix, Arizona, working on agent orchestration: the control planes, guardrails, and evaluation discipline that let fleets of AI coding agents ship real software. He is the builder of GBuild, a macOS AI build environment with a Rust core, a Cloudflare edge backend, a signed app marketplace, and an on-device model line — alongside shipped work in real-time mapping, native iOS and macOS apps, and voice products. His stack runs from TypeScript, React, and React Native through Rust, Swift, and Python. Before software he worked in institutional finance, co-invented a patented mobile sports platform (US 9569922B1), and operated a fitness studio. The through-line is a working rule: verify before claiming, instrument reality, and ship with discipline.
His work sits in the gap most agentic-AI tooling ignores: not making one model smarter, but making fleets of AI coding agents operable — visible, governable, affordable, and safe to leave running. He names the problem first (he calls it the Operator Problem), then builds the control plane the name implies.
02 // The work
One build environment, several systems deep.
The current answer is GBuild (codename Mother), a macOS AI build environment Jeff has built solo since April 2026: 6,308 commits and 275 notarized, auto-updating releases in its first 81 days. An Electron front end rides a Rust core that registers 916 JSON-RPC methods, drives 5 first-party agent drivers (Claude Code, OpenClaw, Hermes, Codex, and a plain shell, plus sandboxed bring-your-own registration), and runs voice, retrieval, and dictation on-device.
Inside it live systems that would be standalone products elsewhere. ThingStore is a signed marketplace for AI-built apps: Ed25519 publisher signing over a SHA-256 content-hash chain, a five-layer automated publish gate, per-app sandboxing down to a loopback egress broker that resolves DNS internally to block rebinding, and revocation that holds offline. The Foundry is an on-device model line — encoder classifiers fine-tuned on synthetic data generated by typed-field corruption, promoted only past paired-bootstrap significance gates on human-labeled gold sets. GBuild Cloud is a 14-Worker (in repo) Cloudflare edge backend: identity built from primitives with no auth SDKs, a key-custody LLM proxy with per-account spend ceilings, telephony with cryptographically verified webhooks, and an autonomy engine governed by off-switches, cooldowns, and circuit breakers.
Before GBuild, the same pattern at smaller scale: PrayerMap, a six-platform React Native prayer app live on the App Store; LivCut, a privacy-first video editor running its AI entirely on-device; and a family of agent-orchestration frameworks whose lessons — file leases, policy-as-code, absence-based alerting — fed directly into the flagship.
03 // The discipline
Verify before claiming.
The habit that organizes all of it is verification. Jeff commissioned a red-team audit against his own portfolio, published the result (most of the site's technical claims initially failed a check against their source repos), and rebuilt the site so every number lives in one canonical ledger with a fact-check script that fails his own build on drift. When his strongest model scored 97.5% on its own synthetic holdout and 62% on a hand-labeled gold set, the number that shipped was the 62 — and the model didn't ship at all, because the promotion gates are code, not judgment calls. The essays on his writing site work the same way: the failure boundary is printed next to the win.
04 // Before, and where
Phoenix, by way of finance.
Before software, Jeff spent years in institutional finance at Morgan Stanley and Raymond James, co-founded FourPlay Football — where he co-invented a patented mobile sports platform (US 9569922B1) — and owned and operated a fitness studio for seven years. The arc explains the engineering temperament: quantify the risk, respect the person on the other side of the product, and treat discipline as the feature that makes the rest possible.
Phoenix, AZ (metro area). Remote-first, open to hybrid or on-site for the right role. Open to work.