Mental health challenges, including stress, anxiety, and emotional distress, are increasingly visible among social media users who share their thoughts and activities online. This paper presents an AI-based framework for detecting stress and suicidal ideation by analyzing multimodal social media activity, including posts, likes, shares, comments, and temporal engagement patterns. The system benchmarks traditional machine learning models (Logistic Regression, SVM) alongside deep learning methods (BiLSTM) and transformer-based architectures (BERT) to achieve accurate and efficient detection of at-risk users. Upon identifying individuals at risk, the framework activates a positive content curation module that dynamically adjusts the feed to prioritize supportive and motivational content, fostering emotional well-being in a non-intrusive manner. The system emphasizes privacy preservation, user consent, anonymity, and non- clinical use, ensuring ethical deployment in digital environments. Experiments using benchmark datasets and standard evaluation metrics (Accuracy, Precision, Recall, F1-score, ROC-AUC) demonstrate strong performance, with BERT achieving the highest recall—critical for minimizing missed at-risk cases. Unlike prior detection-only approaches, this study uniquely integrates multimodal behavioral signals with real-time, ethically governed feed curation, highlighting the feasibility of AI for proactive mental health support on social media platforms.