Things I’m building and learning
A running notebook about building safer AI-agent systems, operating a self-hosted homelab, and turning experiments into practical tools.
I write down what I built, why I chose it, the tradeoffs, and what I learned.
Project notes
Running Hermes like an isolated personal VPS
An always-on personal agent running inside a hardened Ubuntu VM and Docker, with Telegram voice, scheduled automation, persistent memory, rollback checkpoints, and a dashboard designed to stay behind an SSH tunnel.
Jev as a decision gate for agent actions
A typed-probability gate that classifies command risk, estimates blast radius, and decides whether Hermes should run, ask, or block before a tool call. Work in progress.
Turning agent-security ideas into a working queue
A Hermes-managed Kanban covering model routing, alert triage, SSVC and BOD 26-04 decisions, prompt quality, policy drift, dashboards, and personal projects.
Validated diagrams instead of hopeful diagrams
Why I adopted Archify for architecture and workflow diagrams: typed JSON, strict validation, deterministic output, and interactive HTML that refuses broken layouts.
Paperclip and the case for a meta harness
Research notes on coordinating Hermes, Codex, Claude Code, and other agents through explicit tasks, decision gates, traceable artifacts, and per-agent access.
Giving Hermes a practical operating rhythm
OpenCode delegation, a cost-conscious model policy, voice, scheduled digests, vault-integrity checks, and native dashboard deliverables.