AI Electricity Theft Detection

I need an end-to-end workflow that for detecting non-technical losses, addressing class imbalance using SMOTE or cost-sensitive learning from AMI data. My raw information comes from smart meters and customer billing records, stored today as large CSV dumps that ultimately have to live in PostgreSQL. The signals that matter most are usage patterns, any evidence of meter tampering, and billing discrepancies.

Here is the flow I have in mind. You start in Jupyter Notebooks, explore and clean the data, then build and compare Random Forest, SVM and a lightweight neural network so I can see which model produces the most reliable lift comparing accuracy, precision, recall, F1-score, confusion matrix, and ROC curves.

Once the champion model is chosen, wrap it in a small Flask API for real-time scoring. A simple React dashboard should display the theft probability, key feature contributions and allow me to filter by account, date range or location.

Please keep everything in Python (pandas, scikit-learn, TensorFlow/PyTorch—whichever you prefer) and make all queries database-agnostic so they run cleanly against PostgreSQL. The final repo should include:
• Well-commented notebooks that reproduce the exploratory analysis, feature engineering and model training
• The trained model artefacts plus a script to retrain with fresh data
• Flask service with two endpoints: /score (single account) and /batch (CSV upload)
• React front-end served separately or via the Flask app’s static folder
• A README with setup commands and a one-line Docker build so I can spin the stack up locally

I’ll verify delivery by running the notebooks end-to-end on a sample dataset, checking that the API returns consistent probabilities, and confirming the dashboard renders correctly in Chrome.