This paper presents an innovative approach to improving visible light communication (VLC) systems in total shadowing conditions by integrating intelligent reflecting surfaces (IRSs), angle diversity receivers (ADRs), and recurrent neural networks (RNNs). Two ADR configurations (pyramidal and hemispherical) are evaluated, along with signal combination mechanisms: maximum ratio combining (MRC) and select best combining (SBC). The RNN is employed to dynamically optimize the IRS placement, maximizing the signal-to-noise ratio (SNR) at the ADRs and enhancing overall system performance in non-line-of-sight (NLoS) scenarios. This study investigates the spatial distribution of SNRs in VLC systems using RNN-optimized IRSs, comparing the performance of different ADR configurations and signal combination methods. The results demonstrate significant improvements in received power and the SNR compared to non-optimized setups, showcasing the effectiveness of RNN-based optimization for robust signal reception. This article highlights the potential of machine learning in enhancing VLC technology, offering a practical solution for real-world indoor applications. The findings emphasize the importance of adaptive IRS placement and spur further exploration of advanced algorithms and ADR designs to address challenges in complex indoor environments.