Production Model Monitoring

Continuously evaluate machine learning model performance to resolve issues and improve models.

Data analysis

# monitor data for validity, anomalies and drift
sess.log_prediction(
  features={'feat1': 1.2, 'feat2': 'XX'}, output=True)

# compute and monitor model performance metrics
sess.log_evaluation(
  prediction=False, label=True, segments=['state_CA'])

Model Performance Monitoring

Ensure model performance in production and reduce troubleshooting time.

Data monitoring: Monitor offline and online predictions for data validity and anomalies, data drift, concept drift, and more.
Model performance monitoring: Monitor model performance for binary, categorical and numeric models and data segments.
Automatic issue detection: Graphsignal automatically detects and notifies on issues in data and models, no need to manually setup and maintain complex alerting rules.
Any scale and data size: Graphsignal only sends data statistics allowing it to scale with your application and data.
Model framework and deployment agnostic: Monitor models serving online, in streaming apps, accessed via APIs or offline, running batch predictions.
Data privacy: No raw data is sent to Graphsignal cloud, only data statistics and metadata. On-premise version is available for full privacy.

Technology and platform agnostic model monitoring

Python NumPy Pandas TensorFlow Keras PyTorch Scikit Learn Kubeflow Amazon SageMaker Google AI Platform Azure ML Seldon BentoML