Data Scientist,| Software & AI Engineer,| Product Owner

Computes a dwelling's social-housing points score and the maximum legal rent that follows from it, implementing the Huurcommissie policy books against the VERA sector data standard.

Passed Microsoft's exam on designing and building AI solutions on Azure: generative AI apps and agents, computer vision, natural language and document intelligence.

Owns the product backlog, works across teams, and is the first point of contact for the people inside and outside the organization who depend on the microservices.

Fine-tuning a large language model to predict a dwelling's rental points, treating the points score as a regression task.

API monitoring with Prometheus and Grafana across the team's services.

CI/CD pipelines in Azure DevOps, from the first maintenance-team deployments through to the Microservices Team's release process.

Containerised APIs in FastAPI and Docker for the Microservices Team.

Data engineering on Databricks — the pipelines the housing and maintenance data reaches its analysts through.

Co-developed the assessments Woonstad Rotterdam hires data scientists and data engineers with.

Co-founded Woonstad Rotterdam's data science and engineering work group.

DevContainers as the team's development environment, so a new repository is one clone away from running.

Terraform for infrastructure-as-code deployments.

Building, fine-tuning and deploying large language models on Microsoft Foundry, as part of the generative-AI work group.

Exposes the heritage lookup as an MCP server, so a language model can resolve an address to a building and ask whether it is a listed monument.

Passed Microsoft's exam on AI workloads on Azure: machine learning, computer vision, natural language processing, document intelligence and generative AI.

Passed Microsoft's exam on core data concepts on Azure: relational and non-relational data, and analytics workloads.

Looks up the heritage status of Dutch addresses against the Cultural Heritage Agency and Kadaster APIs — tens of thousands of objects at a time, asynchronously, and without an API key.

The heritage status of every addressable unit in the Dutch address register, resolved against the Cultural Heritage Agency and the Kadaster by a staged pipeline built on the monumenten package and published as one Parquet file, with the register behind each status named beside it.

A Streamlit app on Hugging Face Spaces that looks up a Dutch address and says whether it is a national monument, stands in a protected townscape or is a municipal monument, naming the register behind each answer.

Data-quality tests for PySpark DataFrames, with results returned as a long-format DataFrame.

Data-science research on maintenance data, widened to the whole real-estate portfolio on moving to the real estate and maintenance team.

Streamlit applications shipped in Docker containers, putting models in front of the people who act on them.

Checks whether housing data conforms to the VERA sector standard, as a set of tests that plug into the PySpark test framework.

How and why we built a data-quality test framework for PySpark, and what we learned releasing it as open source.

The Huurcommissie's January 2025 changes to its policy book for the rental-points system, for non-self-contained dwellings. Written in Dutch.

The Huurcommissie's January 2025 changes to its policy book for the rental-points system, for self-contained dwellings. Written in Dutch.

A QLoRA fine-tune of Llama 3.1 8B, quantised to four bits, that reads a dwelling's description and answers with its woningwaardering points — regression handed to a language model rather than to a regressor.

A gradient-boosted regressor predicting a dwelling's woningwaardering points from six property features, tuned by randomised search under cross-validation and published with the predictor that loads it, so a dwelling can be scored without retraining.

A Gradio app on Hugging Face Spaces that runs the gradient-boosted model live: enter a dwelling's features and it returns the points it predicts, an interval around them, the maximum legal rent those points allow and the rent sector they put the dwelling in.

Descriptions of self-contained Rotterdam dwellings in a fixed format, each paired with the woningwaardering points Dutch law awards it — generated from the woningwaardering package's own evaluations and split for training.

Dwelling valuations from three housing corporations in one dataset — a collaboration set up between Beveland Wonen, Thuisvester and Woonstad Rotterdam, each publishing its own region — deduplicated per organisation and combined into shared train, validation and test splits.

Tools the standard data-science stack leaves out — plotting helpers, decorators and an estimator switch for scikit-learn pipelines — published on PyPI with documentation on Read the Docs.

A cross-validator for grouped time series where the groups hold unequal numbers of observations — contributed to scikit-lego and shipped in its 0.6.14 release.

A team competition in dynamic pricing: models that set grocery prices in response to demand, competing against the other teams' models.

Bayesian models in PyMC3, where the uncertainty around a prediction is part of the answer rather than something left out of it.

Introduces GroupTimeSeriesSplit, the scikit-lego cross-validator for grouped time series whose groups hold unequal numbers of observations.

Deploying Apache Airflow on Kubernetes, with its scheduler, metadata database and jobs running as resources of the cluster.

Building CI/CD pipelines that test, package and release code on every change, so a merge is the whole of a release.

Two ways to get more out of a categorical plot — counts beside the x-axis labels, and a strip plot laid over a box plot — both from extra-datascience-tools.

Setting up a Python package project with automated tests, documentation on Read the Docs and publishing to PyPI from the first commit.

Building image classifiers in PyTorch and applying them to new images, from preparing the training data to judging where the model fails.

Explaining what a trained model does — partial dependence plots, SHAP values and LIME — so its predictions can be questioned rather than taken on trust.

Four cases in which the best model a GridSearchCV reports is not the best model to use.

Instrumenting APIs with Prometheus and reading the metrics back in Grafana dashboards, so a slow or failing service shows up before its users notice.

Building installable Python packages and shipping each as a Docker container that runs it, so the code is used the same way wherever it lands.

Grid-searching over whole pipeline steps, not only their parameters, with the EstimatorSwitch from extra-datascience-tools.

Deploying cloud infrastructure to Google Cloud as code with Terraform, so an environment can be rebuilt from the repository alone.

A decorator from extra-datascience-tools that logs a function's arguments and running time, set against the icecream debugging library.

Certified in the foundations of Agile and Scrum: the Scrum roles, events and artifacts, and how an agile team plans and delivers its work.

Completed the government's two-year data traineeship: three eight-month Rotations at three government organisations, alongside more than thirty training days in hard and soft skills.

Ran the tutorials for Introduction to Programming, a first course in Python, helping students build their programming skills.

Checked and marked students' Python programming assignments for Introduction to Programming.

Master thesis on telling which artist wrote a song's lyrics. Combined the classic authorship features with part-of-speech, phonetic and statistical ones; questioned whether transcribed punctuation can be trusted as evidence; and showed that adding only an artist's own verses from collaborations beats adding whole songs. The final support vector machines reached macro-F1 scores of 0.91 and 0.79 on two eight-artist data sets.

Graduated cum laude from the Master in Information Science at the University of Groningen.

Ran the tutorials for Advanced Programming, a Python course, helping students take their programming skills further.

Checked and marked students' Python programming assignments for Advanced Programming.

Bachelor thesis classifying Twitter users to one of four parties in the 2017 Dutch general election — CDA, D66, GroenLinks or PVV — by the words they use, the accounts they follow and their hashtags. Collected tweets and followed accounts for 4,000 users through the Twitter API, then compared Naive Bayes with support vector machines, reaching up to 90% accuracy against a 25% baseline.

Set up machine-learning anomaly detectors in Kibana, on data held in Elasticsearch.

Gave a tutorial on how to set up machine-learning anomaly detectors in Kibana.

Ran the store's shifts for a team of up to 15, and handled the cash office.

Trained the new staff who joined the store's team.

Ran the store's shifts for a team of up to 15, and handled the cash office.

Trained the new staff who joined the store's team.

Five Python idioms that mark experienced code, shown through Advent of Code puzzles.

Six practices that make Python code read as written by an experienced developer.

Introduces AdvancedValueCounts, a package that extends pandas' value_counts() for exploring categorical variables.

Cloud gaming on a GPU instance on AWS, paid for by the hour, including Windows games from a Mac with Apple Silicon.

Deploying a Streamlit app that depends on a private GitHub package to Streamlit Cloud.

Getting past junior level in Python by reading, and learning from, the code of professional projects.

Setting up a MacBook for data science work in 2024.

The 2025 edition of the data-science MacBook setup: browser, apps, terminal, Python and AI tooling.

An experiment in whether ChatGPT can write unbiased tests for code it wrote itself.

Whether ChatGPT Plus, and GPT-4 with it, is worth $20 a month to a developer as a coding assistant.

Why git switch and git restore should replace the overloaded git checkout.

Refactoring Emily Bache's Tennis kata step by step in Python, applying the SOLID principles.

Four skills a data science job asks for that a university degree does not teach.