Integrating Large Language Models (LLMs) into Operational Technology (OT) environments presents two fundamental barriers: strict data sovereignty requirements imposed by zero-trust industrial networks, and systematic degradation of LLM reasoning quality when processing raw continuous process data. We present QID-NPI (QID Neural Process Intelligence), a hybrid neuro-symbolic architecture we term an edge-native cognitive hypervisor, deployed in a 2026 industrial dairy and cheese manufacturing facility in Aguascalientes, México. QID-NPI addresses both barriers through a deterministic context preprocessing layer and a multi-agent inference architecture deploying specialized quantized foundation models (including Llama-3.3-70B and DeepSeek-R1-32B) at 4-bit precision on edge hardware. Operating entirely within the plant network boundary, the system generates differentiated outputs per production batch with no data egress. A secondary air-gapped delivery mechanism encodes sanitized prompts into QR codes for local inference on mobile devices, addressing last-mile operator access. Aligning with LLMOps production principles, we demonstrate that transitioning from dense tabular data to a Sparse JSON serialization reduces context payload by up to 77%, significantly reducing Time to First Token (TTFT) while eliminating the Lost in the Middle attention degradation. We further document the commercial and regulatory alignment of this architecture with the December 2025 CISA/NSA joint guidance on secure AI integration in OT environments.
Gomez et al. (Sun,) studied this question.