This series is about the shift from choosing a model to running an AI setup as infrastructure.
The model still matters. But the useful system around it is bigger: agent workflows, local context, memory, tool access, observability, secrets, and boundaries.
Posts in the series
-
From Chatbot to Self-Hosted AI Infrastructure
The overview: why the setup became a stack instead of a chat window. -
The agent layer: the model was never the system
How Hermes, model routing, tool use, scheduled work, Telegram, and Claude review fit together. -
The context layer: context is a system, not a scrapbook
How Obsidian, local retrieval, Qdrant, and memory make the assistant work from sources. -
The boundary layer: tools need limits
Why secrets, Home Assistant, voice, MCP permissions, and monitoring need safe defaults.
Main components
- Hermes for agent workflows and tool use.
- Obsidian as the human knowledge base.
- Qdrant and local retrieval for cited context.
- Honcho-style memory for durable preferences and observations.
- MCP tools for controlled access to services.
- Monitoring and logs for operational visibility.
I am keeping the overview posts readable. Deeper configs and diagrams belong in separate reference posts.