Stock markets are complex, dynamic, and highly volatile systems influenced by macroeconomic indicators, corporate performance, geopolitical events, and investor psychology. Conventional stock forecasting approaches rely heavily on single predictive models, static technical indicators, or human intuition, which are inadequate in capturing non-linear dependencies, regime shifts, and predictive uncertainty inherent in financial time-series data. These limitations increase investment risk and reduce the reliability of automated trading systems, particularly for retail investors in emerging markets such as the National Stock Exchange (NSE) of India. This paper proposes an AI-driven NSE Stock Forecasting and Risk-Aware Trading Decision Support System that integrates classical machine learning, deep learning, market regime detection, and probabilistic uncertainty estimation within a unified multi-model framework. The system employs Linear Regression, Long Short-Term Memory (LSTM), Bidirectional Gated Recurrent Units (GRU), and Temporal Fusion Transformer (TFT) models for multi-horizon forecasting over one to fourteen days. A market regime detection module classifies market conditions into Bull, Bear, or Sideways states and dynamically adjusts model weights in a regime-aware ensemble mechanism, while Monte Carlo Dropout is utilized to generate ninety-five percent confidence intervals to support risk-aware decision-making. A prototype implementation is developed using Python, TensorFlow/Keras, Scikit-learn, Pandas, and Streamlit, operating on historical NSE OHLCV data enriched with thirty-two technical indicators. Experimental results demonstrate that the proposed ensemble framework outperforms single-model baselines in terms of prediction accuracy, variance reduction, and trading signal reliability. The system delivers interpretable forecasts, confidence bands, and automated BUY, SELL, or HOLD recommendations through an interactive dashboard, making it suitable for investors, traders, analysts, and researchers.