Warre Verlinde

Building on Axxes' internal data lakehouse on Azure: new components on top of an existing medallion architecture, with the pipeline work centred on Apache Airflow — designing and maintaining the DAGs that orchestrate ingestion and processing.

The project is in active build-out, and the point of it is what comes after: a foundation that the company's other data initiatives can be built on without each of them solving ingestion again.

On placement via Axxes, building the data environment KBC uses to answer regulatory questions — including ECB onsite inspections, where a supervisor asks something specific and the answer has to be produced from the bank's own data on a short deadline.

Most reporting platforms are built for fixed pipelines on a schedule. This one has to be flexible by design: the questions arrive without warning, they touch different parts of the bank each time, and the answer still has to be reproducible and defensible afterwards.

• Building the environment on Databricks and PySpark, spanning heterogeneous source systems across the bank
• Designing around access and governance rather than bolting it on — with regulated banking data, who may see what is part of the architecture
• Optimizing for turnaround: the bottleneck is how quickly a new question can be answered correctly, not how quickly a fixed job runs

Two parallel client projects in industrial automation, stationed at Orise via Axxes.

Industrial EMS — e-boiler dispatch at Pfizer. Combined energy market knowledge with data science to design a dispatch strategy that is both correct and profitable, then built an economic validation framework so the financial impact of the control logic could be substantiated rather than asserted. I also initiated a cross-validation effort with the client's external energy partner: partly to catch errors, partly because a control strategy that moves money has to be trusted before anyone switches it on.

Advisory system — JDE coffee roaster and freeze dryer. Three-person team further developing an advisory system for industrial process optimization.

• Reworked the data pipeline together with the data engineer, which lifted both data quality and model performance
• Developed a Generalized Additive Model for the freeze dryer, extended with EMA-based deviation correction to absorb process noise that cannot be modelled
• Stabilized the model output through a fine-tuned optimizer, so the advice the operators see stops jumping around
• Improved how raw coffee bean input data is processed for the roaster, which improved its predictions
• Drove the project into its final phase — stabilization, cleanup, and the groundwork to scale to other factories with a proper MLOps and testing workflow

Owned both ends of the same chain: the machine-to-cloud data pipeline, and the forecasting models the Energy Management System runs on.

On the pipeline side, that meant establishing reliable communication between customers' industrial machines, a local edge device, and the cloud — and then owning those flows: designing, implementing and maintaining custom time series pipelines that keep delivering when a machine, a network or a site does not cooperate.

On the model side, price forecasting, energy consumption forecasting and the related predictive models the EMS uses to decide when a customer should buy, store or use energy. These ran in production against real energy bills, which makes reliability worth more than a marginally better score.

DestinE (ECMWF). As part of ECMWF's DestinE project, worked on a Dynamic Line Rating pilot using a local-area weather model — exploring how much extra grid capacity DLR could unlock without putting stability at risk, and analysing what it means to lean on weather predictions for that decision. Largely stakeholder work, with ECMWF, the RMI and Elia.

Sole owner of a computer vision service that detects scaffolding in satellite imagery, taken from "there is no dataset for this" to a monitored service in production.

• Built the dataset from scratch: collected and prepared satellite imagery for a class no off-the-shelf dataset covers
• Designed, trained and fine-tuned deep learning models for object detection and segmentation against it
• Shipped the service to production, including the monitoring needed to notice when it starts to drift
• Analysed model performance and iterated on accuracy and efficiency
• Worked directly with the client to turn a business requirement into a technical one — as the only engineer on the project

Joined as an intern and stayed on as a student worker alongside my studies, working on the design and development of Prophesea's weather forecasting model — the forecasting foundation the energy products were later built on, and the reason I moved into the full-time role.

Seven years at the municipal play scheme (speelplein): preparing and running activities for age groups from toddlers to secondary school students, and coaching and evaluating the animators who ran them. Member of the steering group that shaped the programme.

Applied Informatics with a specialization in Artificial Intelligence, entered through the schakelprogramma from the HOWEST bachelor. The AI track is where the machine learning and modelling work I do now comes from.

Bachelor's degree in Applied Informatics — the software engineering half of what I do, before specializing in AI at KU Leuven.

Secondary education, science-technology track.