DIY legacy AI server tests limits

The idea is not dumb, but the architecture is ambitious in the places that tend to fail first: cabling, thermals, bandwidth, backups, and long-term maintainability.

• –Dual 3090s give the build a real shot at 70B-class quantized inference, but USB4 for the second GPU is the obvious weak link versus a proper PCIe workstation or server board.
• –Markdown and JSON are sane long-term RAG formats; the bigger issue is durable metadata, provenance, embedding regeneration, and regular export tests.
• –LUKS plus physical vault detachment is a good privacy instinct, but legacy access needs documented key custody, threat modeling, and recovery drills.
• –The “personality” goal should probably be RAG-first, not fine-tuning-first, because personal archive grounding matters more than baking fragile behavior into weights.