Past Projects

Project Description: This project documents and operationalizes an end-to-end forecasting system designed to bridge applied forecasting research and production constraints in SMEs. The system is organized in five modules: business scoping, robust data pipelines, modeling, decision optimization, and deployment.

The deployment stack separates control, compute, and state layers. The user dashboard and orchestration trigger batch workflows, specialized microservices execute preprocessing/forecasting/optimization, and storage services handle intermediate artifacts and final decision outputs for partner-specific workflows.

Core engineering elements include meal-level forecasting with deterministic translation to ingredient-level replenishment, probabilistic inference for service-level-aware decisions, and simulation-based optimization under operational constraints.

To ensure long-term continuity and support for client operations, we are currently migrating parts of the platform to an on-premise architecture while preserving the same forecasting and decision-support principles.

Project Description: In this project, we partnered with organizations in the ready-to-eat meal sector in Flanders, including hospitals, school caterers, and meal distributors. The objective was to reduce food waste through improved demand forecasting while increasing each partner's digital maturity for data-driven planning.

We implemented a full forecasting workflow per partner: scoping interviews, data curation, model development, and operational delivery of weekly decision support. Across five distinct partners, forecasting accuracy improved by 13% to more than 133% versus previous planning baselines.

Simulation studies showed up to 67% reduction in food waste per meal and estimated operational cost reductions between 4% and 20%. Four partners actively tested and used the deployed system in weekly production and reordering decisions.

The technical framework and results were presented at the International Symposium on Forecasting (ISF 2025), with follow-up publications in preparation.

Project Description: In this project, we address a primary barrier to AI adoption for many companies in Flanders: the lack of sufficient data. We aim to design and implement effective demand forecasting solutions specifically for small and medium-sized enterprises (SMEs) that possess limited historical data.

Our core methodology overcomes data scarcity by using global models that pool information from related time series, effectively increasing the available data for the AI models. Furthermore, we employ hierarchical forecasting techniques, such as temporal and cross-sectional reconciliation, to intelligently combine predictions from different levels of data aggregation (e.g., monthly, quarterly, yearly), ensuring the forecasts are coherent and make optimal use of all available information.

Our results have shown that this approach achieves a significant error reduction of ... over traditional forecasting methods used by these companies.

To facilitate the adoption of these methods, we have developed and released an open-source Python forecasting package that combines the proposed hierarchical and global modeling techniques.

Project Description: In this research project, we developed a learning analytics system to support at-risk freshman nursing students in a physiology course.

We captured data on student learning behavior by tracking their online activity within the Moodle learning management system and administering a student survey. Halfway through the semester, we used this data to train a Gradient Boost AI model to predict final exam outcomes. To make the predictions actionable, we employed SHAP values to interpret the model's reasoning, allowing us to identify specific learning behaviors that negatively impacted a student's predicted score. Based on these personalized insights, we generated and sent targeted prompts to students, encouraging them to engage in more effective study habits—such as participating in online forums if they hadn't already.

The ultimate goal was to measure whether these data-driven prompts, delivered after the midterm, helped improve student performance on their final exams. We have also developed an open-source Python package to anonymize student data in compliance with GDPR regulations.