Multimodal Neurosymbolic Agentic Systems for Cross-Domain Intelligence Derivation

Abstract

Modern organizations face an urgent paradox: AI systems grow increasingly capable yet progressively less interpretable - and increasingly dependent on external infrastructure that raises legitimate data sovereignty concerns. This paper addresses both tensions by presenting a multimodal neuro-symbolic agentic architecture that treats structured knowledge as the primary engine of intelligence derivation, evaluated across three high-stakes domains - cybersecurity, financial forecasting, and medical imaging - where explainability and privacy are functional requirements. The framework unifies sub-symbolic learning with expert-encoded formal logic through an agentic orchestration layer. All models run locally, ensuring sensitive organizational data never leaves the organization's own infrastructure. In the cybersecurity domain, LTN-induced rules govern malware classification across 24 families from the Malimg dataset, with Redis storing Prolog-encoded logical rules alongside official Microsoft threat documentation. In the financial domain, expert-curated Prolog rules encode market patterns, risk indicators, and sector heuristics, with Redis retaining both formal logic and natural language counterparts. In the medical imaging domain, a CNN+LTN classifier trained on the COVID-19 Radiography Database produces both a diagnostic classification and the symbolic predicates justifying it. The agentic pipeline enforces two-stage validation: Qwen3:32b, as orchestrator, delegates image analysis to Qwen2.5-VL:7b, which verifies whether the radiological image is consistent with reported symptoms - only upon confirmed consistency is the neurosymbolic expert model invoked, otherwise the request is declined. All agents are evaluated via RAGAS across domain-specific query batteries: 24 cybersecurity queries (18 malware-specific, 6 cross-family disambiguation); 30 financial queries (15 highly logical, 15 general); and 30 medical queries (15 genuine COVID-19 cases, 9 valid non-COVID scenarios, 6 deliberately inconsistent inputs). For cybersecurity and financial use cases, an arena-style competition across six local models - DeepSeek-R1:32b, QwQ:32b, Gemma3:27b, Qwen3:32b, GPT-OSS:20b, and Phi4:14b - benchmarks the grounding effect of neuro-symbolic memory on response quality. Results demonstrate that symbolic grounding systematically drives model consensus toward correct classifications, positioning structured organizational knowledge as the active, privacy-preserving backbone through which agentic AI derives trustworthy intelligence - with direct implications for knowledge management in high-stakes learning organizations.