Scottcjn / ram-coffers

What it does

RAM Coffers is LLM infrastructure for inference cost reduction: a NUMA-aware conditional memory architecture that routes model knowledge through physical memory banks, so self-hosted enterprise hardware — refurbished IBM POWER8, Apple Silicon-class machines — serves models locally with no cloud API dependency. It is zero-abstraction LLMOps: the routing layer is the hardware topology itself, measured at 147 tokens/sec (8.8x stock llama.cpp). Verified machines doing this work earn inside the RustChain Proof of Physical AI ecosystem.