We first implement an end-to end system on GNU radio using two USRP X310 for the transmitter and the receiver in an indoor setting. We then build a comprehensive model using the components individually rather than the built-in OFDM transmitter and receiver. We use two different approaches to incorporate machine learning to enhance aspects of the communication system, in a bid to improve channel estimation, in the presence of multi-path fading and frequency offset. Each of our models offers a different advantage. With the first model, JCES performs joint channel estimation and symbol detection. It offers a performance comparable to that of the MMSE estimator even in the case of no cyclic prefix being used. The Cs-CNN (cascaded CNN model) explicitly performs channel estimation and estimates the channel response. This is crucial as it would allow us to further try and maximize the data transfer while keeping the bit error rate as low as possible. Using this model, we hope to use dynamically updating CP-OFDM in the future. Since, synchronization is such an important aspect, we explore a bare bone implementation.