A full infrastructure topology platform covering roughly 90 nodes — Proxmox hosts, k3s nodes, VMs, LXC containers, Docker services, managed switches, all of it — auto-discovered every 60 seconds from eleven different sources without me touching a thing. It figures out the relationships on its own. Which app uses which database. Which service pulls secrets from Infisical. Which pods run on which k3s node. Which apps authenticate through Authentik. Which nodes mount NFS storage from Zhongli. No manual config. It just knows. Click any node and it pulls up status, metrics, and live logs straight from Loki. Pod logs, systemd journals, all of it right there without opening another tool. The click-to-focus mode is the part that makes me happy. Click Nextcloud and it yanks every real dependency out of the graph and rings them around it with typed directional edges — database, storage, secrets, SSO, what node it runs on. You can see exactly what it needs to survive and what breaks if any of those things go away. When something goes wrong, Overwatch suppresses downstream noise. If a host goes down and takes five services with it, you get one alert for the host, not six alerts for the symptoms. It knows the difference. The n8n workflow embeds a unique key in every ticket subject tied to the specific node and metric that fired. Before creating anything it searches existing tickets and compares — if one already exists it reopens it instead of creating a new one. There's a four-path event router handling alerts, suppressed alerts, resolved alerts, and discovery health events separately. A flood guard per node and metric. A two-hour reopen grace window. Orphan handling for suppressed alerts with nothing upstream. When a real alert fires, a lightweight AI router picks one of seven domain specialists — network, Proxmox, Kubernetes, storage, database, security, or incident. That specialist gets the sanitized alert plus live SSH diagnostics from the affected system, and responds with structured evidence-backed JSON that gets posted back as a triage note. If the situation goes stale it re-triages automatically.

Triple-Threat Orchestration System I got tired of AI tools that just do things. Build something, deploy something, write something — and then I find out three steps later that it made a bad assumption in step one that affects everything downstream. Triple-Threat is how I push back on that. It's a structured, gated workflow where three different AI systems each have a specific job, nothing moves forward without passing a gate, and I approve every gate before anything gets built. The coordinator is Hermes — an AI agent with a built-in learning loop that maintains state, enforces the stage sequence, and keeps everyone on task. I don't use it as a general assistant. It runs the workflow. Claude Code does the design and architecture work. Codex handles implementation review and builds the actual runbook. Antigravity (I call it agy) does an independent review of everything Claude Code decided — I make them disagree with each other on purpose, then I decide who's right. Any restart, configuration change, DNS update, secret rotation, or infrastructure action still requires my approval.