Clearbox AI Control Room

Clearbox AI Control Room

An MLops solution to deploy and monitor Machine Learning models in enterprises according to the principles of Trustworthy AI.

AI Control Room

Our solution is model agnostic and can be used to perform a model assessment and to create a production container for model deployment.


Technical Whitepaper

Download our technical whitepaper and find out more about our technology.
For any other information, feel free to contact us.

Model Assessment

The model assessment step can help model owners to identify robustness issues, potential undesired behaviour, and explain errors and uncertainties regarding the model predictions.

The output of such an assessment is:

  • Analysis Report

    A report containing an overview of the most important validation metrics, an analysis of the plausible causes of error and a calibration of the model output.

  • Inference Engine

    An inference model converted into a production ready architecture.

  • Data Augmentation

    The possibility to generate additional training data to further improve the model performance.

Model Deployment

Once a model has been validated our Control Room can be used as an MLOps solution for deploying the validated model. The model is served by creating an augmented API for inference and a database for continuous monitoring. The packaging process is performed using open source and state of the art libraries such as ONNX and FastAPI which guarantee optimal computational performance.

The augmented API is generated by using our proprietary eXplainable AI libraries. This API is designed to:

  • Explanations

    Enrich single queries with local explanations of the output and to enhance the interpretability of the model decisions. Explanations are presented in the form of historical examples, decision rules and counterfactual examples.

  • Monitoring

    Create an intuitive feedback mechanism to monitor the model performance over time.

  • Trust & Reliability

    Perform robust uncertainty quantification of the model output to increase trust in each recommendation.

  • Anomaly Detection

    Perform anomaly detection of live data and report adversarial attacks.