Card default
A random forest and a logistic scorecard on Default of Credit Card Clients (UCI, 30,000 clients). One declined client explained, two models that agree for different reasons, and a feed fault caught as drift.
Six worked examples on public industry data, and runnable notebooks that take a model from first upload to monitored, versioned releases.
The worked examples on this site are ordinary scikit-learn models, trained on public datasets and uploaded with the Python client. Every attribution, counterfactual, insight and drift figure on their pages is what the API returned. Each page names its dataset and says where monitoring traffic was staged from real rows.
A random forest and a logistic scorecard on Default of Credit Card Clients (UCI, 30,000 clients). One declined client explained, two models that agree for different reasons, and a feed fault caught as drift.
Gradient boosting on the 49 Industry Portfolios from the Kenneth R. French Data Library. What drives one month's call, and how the signal holds up in the years after training.
A Poisson gradient-boosting model on freMTPL2freq, 678,013 French motor policies. Rating factors ranked, and one policy's predicted claim frequency explained.
Gradient boosting on Diabetes 130-US Hospitals 1999-2008 (UCI), with patients kept apart between training and test. One discharge explained, and a change in case mix flagged as drift.
Gradient boosting on Telco Customer Churn (IBM sample data, 7,043 subscribers). Why one subscriber is at risk, and the smallest change that keeps them.
A two-layer neural network on Power Consumption of Tetouan City (UCI). One hour's forecast explained in megawatts, and the summer shift in temperature that degrades it.
Five notebooks walk through the whole product on one small public dataset, the Titanic passenger list from Kaggle. It is small enough that each notebook runs in about a minute and every output fits on a page. Outputs are committed, and notebooks 01 to 04 need a free account. A sixth notebook explains an image classifier.
The starting point: data cleaning, feature engineering and eight classifiers compared by accuracy.
Why the model decided as it did for one row, the smallest change that would alter the answer, two models that disagree, and what the forest learned overall.
Two weeks of traffic with a shadow model, then a feed that changes one input's unit, caught by drift detection and named by an explanation.
Error analysis with explanations, better features, then v2 shipped as a new version and compared with v1.
XGBoost, explained with its own TreeSHAP values and through a scoring endpoint that Kaimei calls.
A MobileNetV2 classifier uploaded with its photos attached, then asked why on new pictures. Each answer is drawn exactly as the dashboard draws it.