<p>Real-time intent detection in edge environments faces significant challenges due to computational constraints, privacy requirements, and the need for adaptable responses to evolving user commands. Although cloud dependent Large Language Models (LLMs) can offer strong generalization, their latency and energy demands make them impractical for edge deployment. To address these limitations, this paper proposes a Resource Aware Conditional Cascaded (RACC) framework. It integrates Small Language Models (SLMs) for zero-shot intent detection directly from streaming voice data. Distinguished from conventional approaches, RACC utilizes lightweight semantic embeddings to process overlapping speech streams entirely on edge devices, thereby ensuring data privacy and minimizing latency. We evaluate the framework against standard benchmark datasets under varying acoustic conditions. Empirical results demonstrate that RACC achieves competitive zero-shot accuracy while significantly reducing inference latency and memory footprint. These findings confirm that RACC offers a scalable and robust solution for autonomous, resource constrained smart environments without requiring user specific retraining.</p>