In an era where artificial intelligence is redefining social dynamics, this study introduces a culturally grounded soulmate prediction framework tailored for South Asian contexts. Leveraging a multi-task learning (MTL) architecture, the system concurrently predicts eight relationship dimensions—including gender compatibility, age alignment, occupation match, geographical proximity, emotional outcome, and final compatibility rating based on socio-demographic, psychological, and geospatial data. The neural model is trained on a diverse dataset of anonymized relationship profiles collected across nine provinces in Sri Lanka, ensuring representational fairness and cultural sensitivity. A modular web-based system, implemented with HTML/CSS, Flask, and TensorFlow, enables real-time interaction and user-friendly visualization. The model architecture includes shared encoding layers with dropout, batch normalization, and task-specific output heads using sigmoid, softmax, and linear activations. A custom uncertainty-weighted multi-loss function balances classification and regression tasks. Evaluated using stratified 10-fold cross-validation, the model achieves strong performance across all dimensions (e.g., AUC of 0.96 for gender, R2 of 0.91 for compatibility). The system significantly outperforms traditional baselines like Random Forests and SVMs, while adhering to ethical design principles. This work bridges cultural anthropology, relationship psychology, and machine learning, establishing a new benchmark for ethical, transparent, and context-aware AI-driven relationship modeling.