I have continuous sensor streams (temperature, vibration, current, etc.) coming from a set of industrial assets and I need a reliable machine-learning workflow that turns those raw readings into timely maintenance predictions. The goal is clear: forecast when each machine will require attention so downtime is minimized and spare parts can be scheduled in advance. You will receive a historical data dump (CSV and Parquet) plus a live MQTT feed for validation. After exploring and cleaning the data, build a model—traditional algorithms or deep learning, whichever gives the best performance—that outputs a remaining-useful-life score or binary “service-now” flag for every asset at regular intervals. Python with scikit-learn, TensorFlow or PyTorch is preferred so the final solution stays easy to maintain in our existing stack. Deliverables • Notebook or .py script covering preprocessing, feature engineering and training • Trained model file and clear instructions for re-training with fresh data • Lightweight REST or CLI interface that accepts new sensor readings and returns the maintenance prediction in real time • Brief report summarizing feature importance, evaluation metrics and suggested next steps Acceptance criteria • F1-score ≥ 0.85 (or better than current rule-based baseline) on a held-out test set • Code runs end-to-end on Ubuntu 22.04 with a single shell command • All dependencies listed in a requirements.txt or environment.yml Timeline is flexible within reason, but daily progress notes in our shared Git repo are essential. Let’s turn these sensor logs into actionable maintenance intelligence.