NEXUS: A Distributed Architecture of Ternary AI Agents with Native Data Sovereignty for Multi-Site Enterprise Intelligence

We present NEXUS (Network of EXpert Unified Systems), a distributed architecture deploying ternary-quantized language models (BitNet b1.58) as specialized AI agents on commodity edge hardware. Unlike cloud-centric approaches, NEXUS enforces data sovereignty by design: each agent processes data locally, and inter-agent communication occurs exclusively through natural language via NACP (Natural Agent Communication Protocol), a novel protocol combining NL-only message exchange with cryptographic security (Ed25519 signing, X25519 encryption, anti-replay protection). The architecture introduces a three-level hierarchical orchestration where each level receives only natural language syntheses from below, never raw data. Key innovations include: (1) conversational auto-configuration enabling non-technical deployment, (2) predictive caching through collective usage pattern learning across the agent network, (3) self-monitoring infrastructure agents using the same ternary model architecture, and (4) a device abstraction layer allowing LLM agents to interact with physical equipment through natural language. NEXUS targets multi-site enterprises (retail, healthcare, logistics) requiring real-time operational intelligence while maintaining complete data sovereignty. The system achieves 200-2000x energy reduction compared to cloud-based alternatives, operates on hardware costing approximately €500 per site, and scales from single-site SMEs to networks of 500+ locations. Patent pending (INPI, France). --------- Patent pending, INPI France. The NEXUS architecture and NACP protocol are covered by a French patent application filed in February 2026 by the author. This preprint is published as defensive prior art.