Jeffrey Michael Johnson
AI Engineer & Agent Orchestration · Phoenix, Arizona
Design-driven full-stack engineer building agentic AI infrastructure, local and cloud. I ship products to the App Store, and the systems that orchestrate the agents behind them.
Jeffrey Michael Johnson
AI Engineer, Agent Infrastructure & LLM Systems
Evidence, not adjectives
Eval design, MCP lifecycle, cost routing, orchestration, each carried by a number you can check against the repo, including the one that made a model look worse.
The overfit my hand-labeled gold set caught before a model could ship. Publishing the failure is the point.
Catalog, PKCE OAuth, sandbox isolation, signed bundles. GBuild manages the protocol, it does not just call it.
On-device voice and RAG, cloud only where it earns. One routing map, enforced at commit time.
A fleet of live agents, tiered fan-out, every terminal with a guaranteed shutdown path. Knowing when not to spawn.
Selected work
Each of these fed the next.
The through-line is agent orchestration: making autonomous systems produce work you can actually verify, and publishing the failures alongside the wins. The complete index follows below.
01 · Currently building
GBuild
A macOS AI build environment driving a fleet of coding agents behind 5 first-party drivers, with on-device voice and a signed marketplace.
- Rust core
- Edge backend
- Signed marketplace
Stack
- Rust
- TypeScript
- React
- Electron
- Cloudflare Workers
- Python
- Swift
- Claude
02 · Currently building
G Training Center
The harness lab for GBuild's assistant. Edit the persona and the next run reads it, with no rebuild and no release. The soul A/B runs 3 arms, two of them identical, so an overlay has to beat the engine's own noise before it counts as having done anything.
- Null-baseline A/B
- Human grading only
- Loopback only
Stack
- TypeScript
- Rust
- Python
- Claude
03 · Live on the App Store
PrayerMap
A real-time prayer platform spanning mobile, watch and web, with a glassmorphic design system carried straight over from the Figma prototypes.
- Real-time
- watchOS companion
- Mapbox GL
Stack
- React
- Expo
- Supabase
- Mapbox
- Swift
- TypeScript
04 · Currently building
Latch
DNS automation an agent can drive: detect where a domain is hosted, write the records where it can, and verify propagation server-side against several resolvers. 270 test cases in a 4-day build.
- SPF merge, never a second record
- Multi-resolver verify
- Zero-dependency engine
Stack
- Cloudflare Workers
- TypeScript
- Supabase
- Claude
05 · Currently building
anthill
A local code reviewer running on owned hardware, with a frontier model auditing its work. Built to find out whether free inference can hold a quality bar, and to publish the answer when it cannot.
- Local 8B inference
- Adversarial review
- Measures its own precision
Stack
- Python
- SQLite
- Claude
// Building
Agent orchestration,
end to end.
// Stack
What I reach for.
// Let's talk
Let's build something that matters.
Open to full-time, contract, and fractional or advisory work, especially where an AI system has to survive contact with real users.
Remote-first from Phoenix, travels for client-site work, open to hybrid or on-site for the right role. I reply within 24 hours on weekdays.
If you are building something where the guardrails have to hold, let's talk.
// Direct channels
Email[email protected]LinkedIn/in/jeffreymichaeljohnsonGitHub@jeffmichaeljohnson-techSite87n1.com