Model Development Cycle

The Nexus platform supports custom analytical and processing models that can be deployed directly to Edge devices. These models may include anything from AI and machine learning algorithms to rule-based calculations and signal processing logic.

To manage model development effectively, it is recommended to follow a structured and automated Model Development Cycle — a process inspired by MLOps, which extends DevOps principles to include data and model artifacts.

The overall goal of the process is to improve the following:

• Speed: Reduce the time from experimentation to production deployment by automation.

• Reliability: Use automated and reproducible builds with safe rollbacks.

• Quality: Continuously measure model and data quality through automated validation.

• Compliance: Maintain full traceability of data, code and model versions.

Lifecycle

The lifecycle defines a continuous improvement loop for your models, typically automated through CI/CD pipelines. Each step ensures traceability, repeatability, and collaboration between stakeholders.

The lifecycle of the continuous model improvement process contains the following steps:

The process begins when weaknesses in a deployed model are identified, such as low prediction confidence, poor accuracy (model drift), or business KPI deviations.

• Train A data scientist or developer retrieves relevant datasets — such as misclassifications, low-confidence predictions, or manually labeled examples — from the Nexus cloud data store. Model training can take place locally (offline) or in a cloud development environment. The result is a new model version, including updated code, configuration, and metadata, which should be committed to a code repository.

• Package Once code is checked in, automated build pipelines package the model and its dependencies into a containerized format (Docker image). Each model is treated as a module, versioned and reusable across multiple assets or sites.

• Validate To ensure the quality of the model, automated tests should validate performance against predefined cases. If validation passes, the model image is then published to a Docker repository or directly to the Nexus Module Store.

• Deploy From the Nexus Web Portal, a platform operator can deploy the new model version to either one or more Edge devices (canary deployment), by using the Update Center to update the desired Device Configurations and deploying the changes.

• Monitor Once the new model has been deployed, then model must be monitored continuously in respect to health, performance, drift, data quality and for example business KPIs. The Nexus platform provides tools for this in the form of dashboards, alarms and device monitoring capabilities.

MlOps in Nexus

The figure below shows how the model development cycle can be implemented using the Nexus platform. The Model Development Cycle in Nexus enables a complete closed-loop workflow for model innovation, where continuous improvement is based on operational data located in cloud data store.

Module development cycle using Nexus

The automated pipeline integrates with the Nexus platform, by uploading the model to the Module Store, using either the CLI or API. While deployment can be done automatically, it can also be done manually by an operator.

Model performance must be continuously monitored, where the Nexus Platform provides several tools in the Web Portal, like dashboards, alarms and device monitoring, to supports the operator. For detailed model performance analysis, the data stores can be used.