This paper presents Tempest Sense, a real-time cyclone prediction and monitoring system that integrates advanced data streaming technologies, ensemble machine learning, and a scalable distributed architecture to enable early detection and accurate forecasting of tropical cyclones. The system continuously ingests live meteorological data from the National Oceanic and Atmospheric Administration (NOAA) APIs, including wind speed, atmospheric pressure, sea surface temperature, and humidity, and streams it through Apache Kafka for fault-tolerant, low-latency processing. Cyclone formation is identified using a hybrid ensemble of Autoencoder-based anomaly detection and Isolation Forest algorithms, which together achieve a precision of 0.89 and recall of 0.93, reducing false positive rates to 4.7%. Forecasting is performed via a parallel LSTM and Prophet ensemble, yielding a 24-hour track RMSE of 67.3 km and wind speed MAE of 11.8 km/h. A dual-tier storage strategy using Redis for real-time predictions and ClickHouse for historical analytics underpins the system's performance, with end-to-end pipeline latency maintained under 3.2 seconds at P99. A Flutter-based cross-platform application delivers interactive cyclone track maps, intensity heatmaps, and push-based early warning alerts. The system is designed for deployment in disaster management, climate monitoring, and early warning applications.