~/setup
I engineered my whole workflow.
Not a list of apps. A system. Four devices on a private mesh, an always-on server running local AI, and a personal multi-agent OS I built on Claude. I engineered it as force multiplication: it keeps my research and ops moving while I sleep, so I ship at 10x awake.
4 devices · 1 always-on hub · 70 automations · 2 local models
~/mesh
- skills / automations
- 60+
- watchers
- email-triggered · ~10s
- briefings
- nightly health · weekday brief + EOD
- research
- weekly deep-research → RAG vaults
- orchestration
- orchestrator → workers · fail-closed gates
- reliability
- self-monitoring · auto-recovers
- memory
- layered: session → project → vaults
- lineage
- after Karpathy's LLM-OS / LLM-Wiki
The server is the brain: it never sleeps, watches itself, recovers stuck or failed runs automatically, and ships autonomous work only behind deterministic, fail-closed gates. Cheap and private work, voice transcription, voice-note filtering, local LLM tasks, runs on the RTX 3080. Heavy reasoning and multi-agent orchestration run on Claude. Built after Andrej Karpathy's LLM-OS and his LLM-Wiki memory pattern.
~/cadence
whisper · ollama qwen2.5 · cheap + private
orchestration · deep reasoning · workers
A deliberate split, not a meter: the 3080 absorbs the cheap, private, high-volume work so the cloud budget goes to reasoning.
~/stack
- TypeScript
- Python
- SQL
- Next.js 16
- React 19
- Tailwind
- FastAPI
- Postgres / Neon
- Prisma
- LanceDB
- BGE-M3
- Docling
- Vercel
- AWS
- Docker
- Tailscale
- Syncthing
- systemd
- Claude (Opus/Sonnet/Haiku)
- Ollama · gemma3n
- Whisper
- Square
- Vitest
- Playwright
- tmux
- Git
- Claude Code
~/hardware
- server
- ryzen 9 5900x · 32gb · always-on linux
- gpu
- rtx 3080 · whisper + ollama gemma3n
- primary
- m5 macbook pro
- thin client
- m1 macbook pro · carried daily
- mobile
- iphone 17 pro max · iphone 16 pro · control surface