Gareth Price’s AI setup
Gareth Price runs a headless local-AI workstation with custom agent orchestration (Glass Factory) and llama.cpp on an RTX 4000 Ada, treating the model as both coder and sysadmin while routing tasks across cloud and local models for cost and control.
Key Points
- Primary harness is Claude Code (several hundred sessions/month); custom Glass Factory orchestrator routes Linear tickets through an autonomous pipeline (audit → plan → build → QA → promote) with deterministic state-machine control and human gates.
- Local inference runs on a headless HP Z2 Mini G9 (Ubuntu 24.04, 96 GB RAM, NVIDIA RTX 4000 SFF Ada 20 GB) via llama.cpp; current sweet spot is a 4-bit Qwen 3.8 27B vision model at 128K context (~30 tok/s), with faster smaller models (gemma-4-e4b, minicpm-2b) for quick tasks.
- Access pattern: tmux over SSH/mosh through Tailscale; Sunshine + Moonlight for GUI streaming (including 3840×1080 ultrawide for Xreal One Pro AR glasses); M3 MacBook Pro and Samsung S23 DeX as clients.
- Personal n8n skills: arxiv-digest (weekly scored arXiv papers → Slack), china-digest (Chinese AI news translated), daily-cto-briefing (morning prioritization), MagicMirror LLM Summary (dashboard widget on Samsung Frame TV), and Frigate NVR LLM integration for motion-alert interpretation.
- Team skill pack (private) covers ticket-to-PR flow, git worktrees, standup/incident reports, ISO 24495-1 plain-language rewrites, Gherkin→Playwright browser QA with MP4 demo recording, and pr-review against Notion-cached style guides.
- Glass Factory reporting shows token usage, model routing share, stage advancement, and quality scores; QA uses a different lab’s model than the coder and requires screenshot evidence (mocked renders rejected).
- Agent doubles as sysadmin: durable fixes with “why/how” recorded in a local wiki; zsh-quoting bug promoted to global instruction (use bash for loops/splits/arrays). GPU reserved for inference; remote encoding on iGPU/NVENC; nvtop/btop for monitoring.
- Cost breakdown: Claude Max 100/mo, local inference ~20/mo for larger-model experiments; local still at “fun hobby” economics vs. cloud APIs.