The increasing volume of resume applications creates a significant barrier in front of traditional resume filtering approaches, more often leading to inefficiencies, biases, and inhomogeneous candidate evaluations. In this paper, we propose a novel multi-stage architecture for resume filtering automation utilizing large language models (LLMs) in the tasks of classification and hybrid summarization to facilitate efficiency and fairness. The method integrates LLM-driven zero-shot classification, a multi-agent evaluation framework, and a hybrid summarization module combining extractive and abstractive methods. Compared with a simulated HR ground truth set, the rule-based classifier we propose demonstrated a remarkable precision level of 88.77% and a precision level of 78.39%, qualifying it as a candidate filtering instrument with a strong base. On the other hand, a more semantically complex multi-agent system meant for holistic evaluation showed a 27.92% drop in recall because of its conservative consensus mechanism but has superior modularity and interpretability characteristics critical in high-stakes decision-making scenarios. This paper contributes towards the development of automated recruiting by providing a clear and malleable architecture and placing a strong focus on performance metric vs. explainability trade-offs by utilizing artificial intelligence in recruiting applications.