We present an end-to-end deep learning pipeline for classifying the sentiment of Indian political tweets. We compiled approximately 160 000 tweets from the 2019 general elections using the IndianElection19TwitterData corpus [13] and applied distant-supervision heuristics to label each tweet as ‘pro-BJP’ or ‘pro-Congress,’ discarding neutral or ambiguous cases. This yielded a labeled set of ~75K tweets (≈52% pro-BJP). We trained two neural models: (1) a 1D CNN followed by a unidirectional LSTM, and (2) a bidirectional LSTM with a self-attention layer. Both models used 100-dimensional GloVe embeddings [7]. On held-out validation data, the CNN–LSTM achieved ≈99% accuracy and the BiLSTM–Attention ≈96%. Surprisingly, simply averaging their outputs into an ensemble gave only ≈47.5% accuracy, suggesting the models made correlated errors. Finally, we deployed the trained models in a Streamlit “Votelyzer” web app that offers real-time sentiment predictions on user-input tweets or uploaded CSV files. Our results demonstrate that neural networks can effectively learn political sentiment even from noisy, weakly labeled data, and highlight practical challenges (e.g. label noise and ensembling) in this setting.