Catch electricity theft in your AMI data.
From notebook to live scoring dashboard: GridSentinel scores every smart meter for theft risk, explains why each account is flagged and tells your inspectors where to go first.
From raw meter reads to an inspection list.
Ingest
AMI readings and billing records from PostgreSQL or CSV dumps, validated and deduplicated.
Explore & engineer
Reproducible notebooks build the signals: consumption drops, night usage, tamper events, billing gaps.
Train & benchmark
Three models, class imbalance handled, compared on precision, recall, F1 and ROC AUC.
Serve
The champion model behind a Flask API: /score and /batch, with per-feature contributions.
Act
The dashboard ranks accounts, explains each flag and sends inspectors where theft is likely.
3 interfaces, one application
Choose your role to explore the prototype.
A reproducible repo your team can run and retrain.
Every step is versioned and documented, so the model keeps working when your meter fleet and customer behaviour change.
$ docker build -t gridsentinel . && docker run -p 5000:5000 -p 3000:3000 gridsentinel ✔ API ready on :5000 ✔ dashboard ready on :3000 $ curl -X POST localhost:5000/score -d '{"account_id":"ACC-111896"}' { "theft_probability": 0.982, "risk_band": "high", "top_factor": "Deviation from feeder peers" } $ python scripts/retrain.py --since 2026-10-01 → F1 0.723 vs champion 0.709 · promoted run-0412
See which meters to inspect first.
240 accounts scored on realistic sample data.