Steffen L.

Independent MLOps portfolio project for operationalizing a customer churn model for classification and decisioning scenarios.

In addition to the churn probability, the system generates prioritized retention actions based on the expected economic benefit. This transforms a mere model prediction into concrete decision support.

The focus was consciously not on maximizing prediction accuracy, but on building a reproducible and operational ML system. The model serves as a realistic use case to implement central MLOps patterns such as experiment tracking, model registry, deployment, monitoring, retraining, and CI/CD.

Implemented:

The project demonstrates how a notebook prototype can be transformed into a reproducible ML system. Training, model versioning, deployment, and monitoring are clearly separated and largely automated. This makes changes more traceable, deployments less error-prone, and operational issues detectable earlier.

Tech Stack: Python, FastAPI, MLflow, Prefect, Docker, Terraform, GitHub Actions, GCP Cloud Run, Prometheus, Grafana, scikit-learn, and Pandas.

The complete code, including README, architecture, Docker setup, and CI/CD workflow, is available via my GitHub profile.

Independent MLOps portfolio project for operationalizing a sales forecasting model for demand, revenue, and planning scenarios.

The goal was to develop a reproducible process for training, deployment, monitoring, and retraining from a static forecasting model. The project addresses typical challenges of production forecasting systems such as delayed availability of actual values, changing data distributions, and increasing forecast errors.

The focus was consciously not on developing the best possible forecasting model, but on reliably operationalizing the entire ML lifecycle.

Implemented:

The project demonstrates how forecasts can remain controllable even after the initial deployment. Model quality and data changes are monitored, new actual values can be included in the evaluation, and retraining processes can be triggered specifically. This transforms a one-time trained model into a continuously verifiable ML process.

Tech Stack: Python, FastAPI, MLflow, Prefect, Docker, Terraform, GitHub Actions, GCP Cloud Run, scikit-learn, XGBoost, and Pandas.

The complete code, including README, demo simulation, and deployment documentation, is available via my GitHub profile.

Development and validation of data-driven simulation models for complex technical energy systems.

My responsibilities included setting up reproducible analysis and data processing workflows, processing and validating real measurement data, as well as Design of Experiments, parameter optimization, and sensitivity analyses.

Furthermore, I validated the model behavior using real data, documented model assumptions and limitations, and used the results for decision support, system optimization, and scenario analysis.

This experience forms an important foundation for my current work in the ML and MLOps field, particularly in the areas of data quality, reproducibility, model validation, and reliable technical workflows.