1. Imbalanced Text ClassificationDesign: a Keras text classifier which reads SCADA alarms and suggests the most relevant engineering group.Algo: used NLP to tokenize, lemmatize and stem messages.Performance:I. Accuracy: 82%.II. Speed: 1s per HTTP request.2. Sparse Data Anomaly Detection (Only 1 Usable Variable)Design: a DBSCAN program to predict air compressors’ malfunction and send proactive alarm.Performance:I. Metrics: 83% precision and 83% recall.II. Timeliness: alarms are 2 weeks to 2 hours earlier than failures.III. Speed: 1.3s per air compressor.3. Time Series RegressionDesign: a LSTM time series regressor to forecast water quality in next 24 hours.Performance:I. MAPE: improves from 40% to 30%.II. Speed: 5s per 24-hour forecast.

Business Intelligence EngineeringDesign: cooperated with IT team to install company’s BI data platform into both PC and WeChat mini programs.Result:(1) Automatic and timely stats tracking, with multidimensional filtering and group-by function.(2) Metrics such as absolute amount, percentages, and quarter-on-quarter growth.