Moving SAGE to Local Models
Migrating SAGE's YouTube classifier from a larger local model to a faster one, including the thinking-model gotcha that broke responses.
Migrating SAGE's YouTube classifier from a larger local model to a faster one, including the thinking-model gotcha that broke responses.
Why I added a dedicated Mac Mini to The Hub, how local inference fits the agent stack, and what the first benchmarks taught me.
A routing model for AI work across always-on agents, terminal agents, background desktop workers, and automation workflows.
The moment The Hub needed an operational dashboard: scheduled tasks, GitHub issues, and a visible work model for AI agents.
The meta post -- how this blog was built in a single session with Astro 5, AI-generated images, automated fact-checking, and Firebase preview deploys.
How a growing homelab system got its own domain, a subdomain strategy, and a Cloudflare Tunnel to tie it all together.
Building a heartbeat system that lets AI agents act on their own schedule — birthday reminders, GitHub scans, TV recommendations, and more.
Building a gateway to expose homelab AI agents to Claude.ai and Claude Desktop, and the outputSchema debugging saga that almost broke everything.
Adding a fourth AI agent to The Hub — a knowledge curator who classifies YouTube videos, recommends TV, and proved that adding agents really is a content problem.
How n8n became the shared automation layer for my AI agents, exposing integrations as MCP tools without agents needing to know about API credentials or plumbing.
Scaling from one AI agent to three on a single LXC container, with agent-to-agent delegation, per-agent MCP servers, and an n8n automation layer.
Building a mobile-first PWA that captures photos, links, and text on the go, then uses AI to enrich and route them into The Hub knowledge base.
Why I replaced Discord with a custom Angular PWA for talking to my AI agents, and the Firebase IAM gotcha that almost derailed the deploy.
How I built an autonomous AI agent on a 2GB LXC container in my homelab, complete with an X-Men persona and a subscription-based LLM strategy.
I built QuickShare -- a Claude Code command that takes any HTML file and deploys it to Firebase Hosting with social-ready OG tags. Because if sharing is hard, you won't share.
I discovered the Dutch basketball federation runs on a public API. So naturally, I reverse-engineered it to build scouting reports for my kids' U12 teams.
From streaming service comparisons to interactive infographics, weekend three of the AI Resolution closed the trust gap between 'AI can research' and 'I trust this for decisions.'
I tested Claude, GPT, and Gemini on identical tasks to build a personal model map. The results surprised me — especially what mattered most.
How a Marvel-themed Obsidian vault, a GitHub connection, and a 10-weekend AI challenge kicked off the system that became The Hub.