I wanted to understand how platforms actually work — not just click buttons on managed K8s. So I built one from scratch. Eight machines, three single-node K3s clusters, VPN mesh, SSO, full observability. Everything IaC, nothing manual.

AI assistants waste tokens loading context they don't need. Hive queries your Obsidian vault on demand via MCP — only what's relevant, when it's relevant. 67-82% token reduction.

Cloud LLM APIs see everything you type. Pollex polishes your English on a $99 Jetson Nano — grammar, coherence, wording. A measured median of 6.4 seconds per request with one user; time grows with the text, about 16 seconds for 350 characters. A cloud engine is optional.

Corporate firewalls block VPN clients that need admin rights. ts-bridge tunnels RDP/SSH through Tailscale in userspace — no install, no admin, no traces. One binary.

Find-and-replace in PDFs without destroying the formatting. CLI for batch jobs, MCP server for AI agents.

YouTube Studio analytics are shallow and locked to your own channel. This CLI pulls metrics for any public channel and exports everything.

Setting up smart plugs through the vendor's phone app ties them to its cloud. This provisions a factory-fresh plug onto your WiFi, finds the plugs on the LAN and switches them locally, over both TP-Link protocols.

Dev environment that rebuilds itself. Zsh, Neovim, tmux, Git — one script on a fresh machine.