Browser-Native Federated Inference on Existing Italian SSN Clinical Workstations: A Peer-to-Peer Sovereign AI Architecture for Italian Regional Health Networks

Background: The clinical adoption of large language models (LLMs) in public healthcare faces a structural impasse. On one side stands the prohibitive capital expenditure of centralised high-performance computing infrastructure; on the other, the substantial privacy risks involved in routing sensitive patient data through third-party cloud APIs. Italian Local Health Authorities (Aziende Sanitarie Locali, ASL) operate hundreds of thousands of clinical workstations that remain idle outside peak administrative hours—an untapped computational reserve that, if put to use, could power local AI inference without any external dependency. Objective: We propose and evaluate OmniMed Federated, a decentralised, browser-native architecture that uses the WebGPU API and the WebLLM framework to build a self-organising peer-to-peer (P2P) network for LLM inference across institutional workstations. The primary goal is absolute data sovereignty: patient-identifiable information must never leave the institutional network perimeter under any operational condition. Unlike established federated approaches that require dedicated edge servers, containerised infrastructure, or native software installation, OmniMed Federated operates entirely within the browser on hardware already deployed in ASL workstation fleets—eliminating procurement overhead and enabling immediate institutional adoption without administrative privilege requirements. Methods: We designed an "Edge-First" orchestration layer that harvests idle compute cycles from existing clinical hardware without requiring software installation, administrative privileges, or infrastructure modification. A five-tier escalation model, managed by a lightweight PHP backend (OmniMed Backend Prototype), coordinates node discovery, task assignment, and graceful fallback. The system is optimised for the hardware homogeneity typical of Italian regional procurement frameworks, which produces predictable performance profiles across ASL workstation fleets. Results: A three-node experimental testbed operating on a shared institutional Wi-Fi network achieved a federated throughput of 19.5 tokens/second—a 137% improvement over single-node execution—while reducing peak per-node VRAM consumption by 62%. Node discovery latency was 140 ms. Patient data residency remained 100% local at all times. The system maintained operational continuity during simulated internet blackout conditions. Conclusions: A federated, browser-based compute-sharing model is technically feasible on existing NHS hardware and offers a scalable, GDPR-compliant pathway to meaningful AI capability in resource-constrained public health environments. The architecture described here is operationally live and openly accessible.