04 / Project
Operator AI System — Overseer Stack
A personal command system that routes voice and iMessage requests, dispatches Mac/VPS agents, tracks approvals, and keeps an audit trail.
The operating problem
AI work gets fragile when it lives in one chat window. Real operations need intake, routing, state, approvals, logs, and a way to hand work between people and agents without losing the thread.
That is what the Overseer stack is for.
What it does today
- Accepts trusted voice and iMessage requests, turns them into command-center jobs, and records a receipt.
- Dispatches work to Claude and Codex panes on the Mac or VPS, then watches for completion, failure, or pending approvals.
- Keeps shared task, memory, health, heartbeat, and audit context available to the next agent instead of burying it in a chat transcript.
- Separates operator approval from execution so the system can move quickly without sending messages, spending money, or touching production by accident.
Architecture at a glance
voice / iMessage / web request
|
v
Overseer intake -> command jobs -> approvals -> Mac + VPS agent panes
| | |
v v v
memory + task state receipts health / audit trailWhy it belongs on a CV
The same failure modes show up in manufacturing AI: unclear handoffs, no audit trail, weak approval gates, and tooling that works only while the builder is watching it. This system is a small, personal version of the operating layer I like building around AI: clear intake, controlled execution, measurable state, and recovery paths when something breaks.