Bastion — Private AI Cloud-in-a-Box on AMD

Enterprises want department-grade AI — legal that answers like counsel, finance like a CFO — but hit a three-way bind: per-token APIs leak sensitive data, a dedicated GPU per fine-tuned model destroys the economics, and one-size models waste budget on trivial queries. Bastion solves all three on a single AMD GPU. A base model plus one LoRA adapter per department are served simultaneously from one card via vLLM multi-LoRA on ROCm; adapters are fine-tuned on the same box, so private data never leaves it. In front sits a policy router that classifies every query in under a millisecond: raw-sensitive content is pinned to local engines (live egress counter reads zero), PII-only queries are automatically redacted and the declassified residual is served by the cheap Fireworks AI serverless tier, hard questions escalate to a larger local model, and everything else hits the department's own adapter. The Bastion Console — a branded, mobile-responsive admin platform — lets teams onboard a new department live: one form hot-loads its adapter into the running GPU with zero restart and zero new hardware. Every routing decision lands in a hash-chained, tamper-evident audit trail with one-click CSV export and chain verification — the compliance artifact regulated buyers actually ask for. Everything shown is real infrastructure built this weekend on AMD Developer Cloud (ROCm 7.2 + vLLM + PyTorch, Radeon PRO W7900). The same code scales unchanged to AMD Instinct MI300X, where 192 GB HBM3 hosts dozens of departments — plus a co-resident 70B hard-tier — on one card. One AMD GPU. Every department. Zero egress.