Neural Architecture Search (NAS) automates the design of neural networks but faces the prohibitive cost of evaluating eachcandidate architecture through full training. Zero-cost proxies estimate architecture quality in seconds without training, yetindividual proxies remain noisy, and no existing method jointly exploits proxy signals and the topological structure of architectureswithin an active-learning framework. We propose ZAPS (Zero-cost Active Proxy Search), a four-component pipeline that addressesthese limitations. ZAPS first selects a compact, non-redundant subset of proxies via PROXYFIT, a greedy anti-redundancy algorithm; initializes the search with a hybrid K-means strategy balancing exploitation and exploration; dynamically refines proxy selection using bootstrapped MRMR; and predicts architecture quality with an XGBoost ensemble fed by proxy ranks and one-hot topological encodings, guided by an Upper Confidence Bound (UCB) acquisition function. On NAS-Bench-201 with a budget ofB =200evaluations, ZAPS identifies 53.8% of the top-100 architectures on CIFAR-10 and 66.8% on CIFAR-100 (P@100), while maintaining a Regret of 0.04% on CIFAR-10. ZAPS outperforms comparable-budget methods — REA, BANANAS, and TPE — on both metrics evaluated on CIFAR-10 and CIFAR-100 of NAS-Bench-201.