Technology has an important role to play in the field of Rehabilitation, improving patient outcomes and reducing healthcare costs. However, existing approaches lack clinical validation, robustness and ease of use. We propose Tele-EvalNet, a novel system consisting of two components: a live feedback model and an overall performance evaluation model. The live feedback model demonstrates feedback on exercise correctness with easy to understand instructions highlighted using color markers. The overall performance evaluation model learns a mapping of joint data to scores, given to the performance by clinicians. The model does this by extracting clinically approved features from joint data. Further, these features are encoded to a lower dimensional space with an autoencoder. A novel multi-scale CNN-LSTM network is proposed to learn a mapping of performance data to the scores by leveraging features extracted at multiple scales. The proposed system shows a high degree of improvement in score predictions and outperforms the state-of-the-art rehabilitation models.