Yagub Rahimov, Vignesh Karumbaya
Defense Applications Lead, Texas, United States
Keywords: Small Language Models, Adaptive Model Orchestration, Explainable AI, Low SWaP AI, Edge AI
Polygraf has developed a novel adaptive architecture that replaces dependence on a single, general purpose large language models (LLMs) with a coordinated network of task-specific small language models (SLMs). Each compact model is optimized for a defined mission function, including sensitive-data detection, content-authenticity analysis, provenance anomaly detection, and policy-based classification. A lightweight orchestration layer routes each input to the appropriate models, fuses their findings, assigns confidence, and refers uncertain results for human review. SLMs enable accurate, explainable AI to operate locally on standard computing hardware, including secure, disconnected, and resource-constrained environments. Unlike cloud-based LLMs, the architecture can function without external API calls, dedicated GPU infrastructure, or the transfer of sensitive data outside the user’s control. Its modular design also allows individual models to be updated for new threats, data types, and mission requirements without retraining the entire system.