A personal lab for building, running, and understanding artificial intelligence — from local inference to multi-model orchestration, one experiment at a time.
Milestones on the journey from concept to always-on assistant
Mac Mini running 24/7 as the central hub. Web chat, iMessage, voice, and Obsidian channels all connected. Local Ollama models handling the baseline.
COMPLETESmartRouter classifies each request and dispatches to the best backend. Claude, Gemini, Grok, and Ollama work as a team. Cost-optimized, capability-maximized.
COMPLETELong-term memory extracts facts from every conversation using local models. Morning briefings, evening summaries, and goal tracking make the assistant proactive.
COMPLETEEmail server at vblabs.ai, landing page, and infrastructure under our own domain. Reducing cloud dependency one service at a time.
COMPLETENVIDIA DGX Spark for local model serving. 128GB unified memory, Blackwell architecture. The goal: handle 90%+ of requests without touching the cloud.
PLANNEDThree pillars of hands-on AI exploration
Running open-source models on personal hardware. Understanding what's possible when your AI lives on your desk, not in someone else's cloud.
Routing requests to the right brain — Claude for reasoning, Ollama for privacy, Gemini for breadth, Grok for speed. One assistant, many minds.
Self-hosted email, voice, messaging, and knowledge management. Building the stack that makes an AI assistant truly yours.
Principles guiding the exploration
Your AI should run on your hardware, store data in your databases, and answer to you — not to a platform's terms of service.
Begin with a working prototype. Add complexity only when you understand why it's needed. A $600 Mac Mini before a $4,000 GPU server.
No single model is best at everything. The real power is in routing — knowing which brain to use for which task, automatically.
Share what works, what doesn't, and what's surprising. The AI landscape moves fast — the best way to learn is to build and document.