A novel multi-layer LSTM-based approach for signal detection and channel estimation for OFDM wireless systems is proposed in this paper. Unlike conventional methods such as Minimum Mean-Square Error (MMSE), Least Squares (LS), and Fully-Connected Neural networks (FCN), the multi-layer LSTM model leverages its capability to capture temporal dependencies and sequential patterns in the received signal, enabling more accurate signal detection and Channel State Information (CSI) estimation in an end-to-end manner. We also evaluate our proposed model through extensive simulations, focusing on its robustness in various scenarios, including varying numbers of pilots, different cyclic prefix (CP) lengths, and a combination of reduced pilots and CP. Furthermore, we examine the impact of mismatches between offline training and online deployment environments, where channel conditions differ between training and real-time operation. Simulation results indicate that our approach consistently achieves superior performance in terms of BER compared to LS, MMSE, and FCN. Our model also shows strong resilience to mismatched deployment environments, maintaining robust performance even when the channel characteristics during deployment deviate from those in training.