Recommender systems are important to provide users with the relevant content by learning through the past interactions. Nevertheless, the current recommendation systems have a problem of popularity bias i.e. such system recommend popular (frequently interacted) items and not long-tail items. This disproportion has a detrimental effect on diversity, equity, and user satisfaction. Even though the methods of reranking and debiasing have been suggested in previous research, not all of them are effective in terms of balancing accuracy and improved exposure to less popular items. In order to fill this gap, the present paper introduces a tail-aware hybrid recommendation system, a combination of Bayesian Personalized Ranking (BPR), feature-enhanced LambdaMART reranking, and maximal marginal relevance (MMR) to complete the final selection. One of the central novelties is the long-tail-sensitive negative sampling approach that complements the prior method of contrastive learning by comparing rare positive examples with popular negative ones with the help of implicit feedback and textual pretrained embeddings. The proposed method, tested on Amazon product data, obtained 87.55% tail item coverage, 0.6439 diversity, and 0.7442 serendipity, which is 47 and 27 times better than existing methods in tail coverage and diversity, respectively, and is as accurate as its competitors (0.6957). These findings indicate that the framework is an effective scalable and interpretable solution to reduce popularity bias in practice in recommendation contexts.