What I use AI for
Apart from work, for my hobby I run a homelab that I treat as a real environment: media services, live TV, the usual self-hosted sprawl. Most of my AI use is coding and infrastructure work on that.
I route and/or track everything through my own self-hosted otari gateway (https://github.com/mozilla-ai/otari), which is OpenAI- and Anthropic-compatible. That lets me put local models and hosted models behind one endpoint and switch per task. Local models do the bulk work. Hosted models do the things local models are bad at: long context, speed-sensitive jobs, anything where I'm going to be waiting on it.
My everyday setup
- Agent of Empires (https://github.com/agent-of-empires/agent-of-empires) for running several agents in parallel in tmux.
- pi (https://pi.dev) as the coding agent, pointed at my own otari gateway rather than a vendor API.
- otari (https://github.com/mozilla-ai/otari) as that gateway: one endpoint in front of local and hosted models.
- DwarfStar (https://github.com/antirez/ds4) running DeepSeek V4 Flash 0731 on a Mac Studio 128GB VRAM
- pi-referee (https://github.com/njbrake/pi-referee), A second cheaper and faster model reviews every tool call against plain-language policies and escalates anything doubtful to me instead of blocking silently.
- Claude Code for the heavier sessions where I want the stronger model.
How I work
YOLO Mode, as a default policy. I run all agents in AoE inside docker containers so that I have confidence about what each agent can and cannot touch.