The document contains abusive and profane language. The increasing prevalence of hostile content on social media has raised significant concerns among governments and organizations, necessitating robust AI-based solutions for effective content moderation. While considerable progress has been made in detecting hostile posts in English, the Indian language domain, particularly Hindi, remains underexplored. Existing research in Hindi is constrained by limited datasets, hindering the development of high-performing models. To address this, we developed a comprehensive multi-class dataset [1] tailored for Hindi, encompassing hostile categories such as defamation, abusive language, hate speech, offensive language, and non-hostile content. Leveraging this dataset, we propose a novel hybrid architecture that integrates statistical and contextual embeddings to enhance classification performance. Statistical embeddings, extracted via Term Frequency-Inverse Document Frequency (TF-IDF), are utilized with five machine learning classifiers, evaluated through 5-fold cross-validation. Contextual embeddings, derived from Multilingual Representations for Indian Languages (MuRIL), are employed in both the deep learning model and the hybrid model. The proposed hybrid model, combining an Long Short Term Memory network with a Random Forest classifier and leveraging contextual embeddings, achieves remarkable performance across multi-class classification tasks, with an accuracy of 0.9892 and an F1-score of 0.9893.