Title : NEXRIVOS: Cross-jurisdictional food safety analysis multi-agent large language model orchestrator with age-adaptive consumer education
Abstract:
Identically branded food products are not compositionally consistent across jurisdictions, ingredients permitted in one region may be banned or restricted in another, producing systematic formulation differences driven by fragmented regulations and heterogeneous supply chains. Consumer food-safety tools, However, these systems operate within a restricted regulatory context and rely on barcode-indexed or product name lookups, rendering them inherently database-dependent and limiting their ability to detect cross jurisdictional inconsistencies or keep pace with evolving ingredient compositions. NEXRIVOS (Nutrition Expert Regulatory Intelligence Verification and Orchestration) reframes food-safety assessment as a cross-jurisdictional, explainable inference task and implements a staged pipeline that separates deterministic regulatory intelligence from constrained generative reasoning. An OCR-first ingestion module (CLAHE, Google Cloud Vision+EasyOCR, domain-specific NER) extracts ingredient lists from unconstrained labels. A deterministic regulatory layer normalizes heterogeneous corpora, resolves entities to standardized chemical identifiers (e.g., PubChemCIDs), and performs low-latency retrieval methodology across FDA, EFSA, FSSAI, GB2760, and FSANZ, with embedding-based fallback for unresolved entities. Multilingual canonicalization (MarianMT) and scoped LLM reasoning handle multi-ingredient interaction hazards, while an age-adaptive educational layer produces developmentally calibrated explanations. Evaluation reveals widespread cross-country formulation divergence, benzene-formation risks involving erythorbic acid (E315), and detection of cyclamates (E952), banned in U.S. foods but permitted elsewhere. Empirical validation demonstrates NEXRIVOS’s ability to detect benzene-formation risk in three commercial beverages, identify a product containing an ingredient banned from U.S. foods from 1969, and quantify a 30% cross-country ingredient discrepancy rate-findings unreachable by single-jurisdiction systems. NEXRIV demonstrates that food safety is inherently cross-jurisdictional and establishes a generalizable design pattern for hybrid, explainable decision-support systems.
