Jeffrey Michael Johnson

Rust (tokio) · TypeScript · Electron · React · Cloudflare Workers/DO/Queues/KV/R2 · Python (ML + voice) · Supabase · Stripe · Telnyx · LiveKit

• Built a notarized, auto-updating macOS AI build environment: 796,773 hand-written LOC across a Rust sidecar (1045 registered JSON-RPC methods, panic-isolated dispatch, 994 source files) and an Electron/React front end, with 344 notarized releases distributed via Cloudflare R2 in the first 121 days
• Designed an 18-Worker (in repo) Cloudflare edge backend (Service Bindings, Durable Objects with SQLite + Hibernation, Queues, KV, R2, Analytics Engine, Cron Triggers) covering auth, telemetry, an autonomy engine, a voice/telephony hub, and a deploy-gated billing rail
• Implemented identity from primitives, no SDKs: magic-link auth, 4-provider OAuth2 with hand-rolled PKCE, a 16-connector token vault with AES-GCM encryption at rest, and dual-token Supabase JWT minting for RLS-scoped Realtime
• Built ThingStore, a signed app marketplace: Ed25519 publisher signing over a SHA-256 content-hash chain, a 5-layer automated publish gate (schema validation, secret/PII scan, capability-ceiling linting, CSP static analysis, LLM risk review), egress-allowlist scrutiny, and a revocation feed that holds offline
• Built The Foundry, an on-device model line: fine-tuned ModernBERT/DeBERTa classifiers on synthetic data generated by typed-field corruption, evaluated against human-labeled gold sets behind paired-bootstrap 95% LCB significance gates with a champion/challenger harness, and shipped a Rust runtime ranker with a versioned feature contract to the Python trainer
• Engineered agent-runtime infrastructure: a pluggable AgentDriver trait behind 5 first-party drivers plus sandboxed bring-your-own-agent registration, coordination groups over a pub/sub bus, advisory file leases with deadlock-avoiding sorted acquisition, and a policy-as-code governor (task class → model tier, cost ceiling, tool least-privilege) enforced by a build-failing conformance test
• Cut per-turn LLM cost by trimming ~96 MCP tool schemas to each agent’s capability set, byte-deterministically so the prompt cache still hits
• Hardened the untrusted-code surface: macOS Seatbelt (sandbox-exec) SBPL profile generation, per-app Electron session partitioning with deny-by-default permission handlers, and a loopback HTTP CONNECT egress broker that defends against DNS rebinding and SSRF
• Built AI-safety operational infrastructure across the agent fleet: off-switches, circuit breakers, cooldowns, an append-only audit trail, and absence-based alerting that treats a missing heartbeat as the signal

OpenAI Realtime · Gemini Live (3.1 Flash Live) · MLX Whisper · Kokoro TTS · Telnyx · TypeScript · Rust

• Architected a voice-first control plane spanning dictation, conversational, and realtime modes (OpenAI Realtime, Gemini Live) with on-device MLX Whisper STT and Kokoro TTS: an intent classifier and router turns speech into tool calls that operate macOS apps, files, and folders, GBuild itself, and the agent fleet, guarded by destructive-intent confirmation and transcript redaction
• Built a custom Voice Harness Lab plus training center for voice agents: real multi-turn conversations against live harnesses where per-turn human feedback becomes an approved patch to the harness source effective on the next run, alongside scripted voice skits, an intent-router bench, consent-gated dataset export, and replayable session journals
• Designed a proactive voice agent pattern (the agent probes for ideas, interests, and goals instead of waiting for prompts) and demonstrated it in live internal sessions completing a PRD and personal and business CRM records in real time from the running transcript
• Programmed AI agents to place real phone calls over Telnyx telephony behind a Cloudflare Workers comms hub; deployed internally in a test environment for internal testers

Python · Qwen3-8B via MLX · Claude Opus · SQLite (WAL) · FastAPI · launchd

• Built an agent factory inside it: a person types a request, a cloud model plans and writes the agent, and it lands in the repository behind a step that names itself the trust boundary, where every build-shaped field the model returns is machine-verified or refused. The safety argument is a single property held by two named tests rather than a promise, that a generated agent is merged but never scheduled, because merging is reversible and scheduling is not. 3 agents written this way are in the repo
• Built a 24/7 code reviewer running an 8B model on a spare Mac mini: 8,826 review runs across 2,480 file versions, 24.88 million prompt tokens at zero marginal cost, with Claude Opus auditing samples of its output
• Made the system publish its own error rate rather than its volume: every figure it reports is regenerated from the live database at render time and the check fails when the page and the database disagree, built after an earlier hand-typed draft overstated its own precision by 3.5x
• Reported the result that argues against the design: the paid audit has upheld 12 of 715 findings, and a verification rebuild that lifted the upstream refute rate from 0 to 37.5 percent left end-to-end precision lower than before it, 4.8 percent down to 1.4 percent
• Put the cheapest filter first: a deterministic, no-model grounding stage rejects findings that quote code absent from the file, discarding 572 of them at zero cost before any billed call

TypeScript · React · Node · installed-app engine against a swappable home directory

• Cut the edit-to-observation loop from a full cloud build to the next run by driving the shipped app's own engine against a swappable home directory, so a persona edit takes effect immediately with no rebuild and no release
• Designed the experiment around a null baseline: three arms, two of them deliberately identical, so a persona overlay has to beat the engine's own run-to-run variance before it counts as an effect. Grading stays human on purpose, since an automated judge would be a second unmeasured layer scoring the first
• 31 API endpoints and roughly 11,000 lines of tooling; the parity manifest the lab reads is generated from the production Rust source so the lab cannot silently drift from the app it is tuning

Next.js · React · TypeScript · Supabase (auth + Postgres) · Resend · Three.js

• Delivered and shipped a production platform for an outside client inside their constraints and on their timeline: 345 commits over six weeks, live in production, with authenticated accounts, transactional email, and a WebGL front end
• Owned the engagement end to end: scope, build, deploy, and handoff. Client name withheld here and available on request

Figma · React Native · Reanimated · Skia · Expo · Supabase · Mapbox · SwiftUI (watchOS) · LiveKit

• Designed the "Ethereal Glass" UI/UX and animation system in Figma (color system, typography, glassmorphism tokens, 8 category-specific glow colors, 6-phase prayer connection choreography), then translated the wireframes and prototypes into a production React web app before pivoting to React Native based on design-fidelity requirements
• Authored a 470-line animation preservation spec mapping every Framer Motion interaction to Reanimated and Skia equivalents; iterated on the front-end implementation through a dedicated animation lab (prayerConnections, memorialHub experiments) to hit a 60fps native target
• Shipped on the App Store across three Apple platforms (iPhone, iPad, Apple Watch), most recently v1.1.0 on 2026-08-28, backed by 49 Supabase edge functions, 287 database migrations, a native SwiftUI watchOS companion with complications, and real-time subscriptions
• Led v1.1.0 on accessibility and privacy rather than features: expanded VoiceOver labelling, plain-language purpose strings on every camera and microphone prompt, no server calls for prayer connections while signed out, and open-source licences readable inside the app
• Currently building the visionOS spatial app (build 4, not yet submitted): a grabbable Living Map you can spin, zoom and fly through, with prayer markers matching the mobile experience
• Built 6-dashboard admin command center with 3-tier AI behavior monitoring, Datadog RUM, and PostHog analytics

Node.js (ESM) · Claude Code CLI · MCP Protocol · ChromaDB · Docker

• The Smart MCP Loader + 13 skills + 7 specialized agents that form the operating system I build inside. NLP-scored prompt routing across 11 adaptive tiers, MIT License.

Node.js (ESM) · Claude Code Hooks · Regex NLP Scoring · node:test

• 183 passing tests, 0 failures. Classifies prompts on a 4x4 complexity/stakes matrix, routing to 11 adaptive tiers. Adversarial-tested with multi-agent parallel validation.

Python · LangGraph · FastAPI · Pinecone · MCP Protocol

• 7 specialized LangGraph agent types, 186 commits of WordPress hook-mapping research. Pinecone vector retrieval with hybrid knowledge graph architecture.