Builds operator-grade systems for one human and many agents.
I design and ship the orchestration, memory, and approval layers that let a single operator run a fleet of LLM-driven agents without losing the thread.
02 / Work
Eight systems in production. Each one earns its place.
Built across the last twelve months. Each project is a tool I use daily, not a one-off for a demo. Click into a row for the dossier.
- 01
Atlas
Live registry and graph of every component, host, service, schedule, queue, and data store across the stack.
- 02
ChatGPT Codex Bridge
HTTP MCP bridge that lets ChatGPT and Claude web call into a private Codex and Overseer surface without exposing path tokens.
- 03
JOB-SNIPER
On-command autonomous job-application system with deterministic claim-diff gates and persona-locked tailoring.
- 04
Memory Pipeline v2
Two-hook capture layer plus a pgvector store that gives every Claude session durable memory and cross-host recall.
- 05
OpJune
Budget gate, raw-caller deny shim, and capability layer that lets the Overseer brain hold real money without losing the kill switch.
- 06
Overseer
Multi-host orchestrator for Claude and Codex panes, with a brain that proposes actions and a public MCP surface for ChatGPT and Claude web.
- 07
Unified Task Engine
Single canonical task queue across Mac and VPS. Replaces a small zoo of launchd, cron, systemd, and ad-hoc Python schedulers.
03 / Conversation
Ask the operator anything.
Real-time voice or text. The answer cites which role or project it pulled from. Voice runs on a polished ElevenLabs voice; cloned-voice slot is reserved until John drops a 30-second sample. Text mode is the always-on accessibility fallback.
04 / Resume mirror
Drop your resume. Get a tailored pitch.
Drop your CV (PDF or text) and choose a direction. A vision-capable GPT either pitches you to me, or pitches me to you — your call. Rate-limited per IP. No data stored.