Rainfall Prediction
Top 0.2%Predicting daily rainfall from weather features. After feature engineering and K-fold cross-validation, a well-tuned KNN beat heavier ensembles. Finished rank 5 of 2,529.
Data Engineer at Rabobank·Leiden, NL
WHERE I'VE WORKED AND STUDIED
Now
Data pipelines and ML infrastructure in production: MLOps, CI/CD, infrastructure as code, automated testing and monitoring.
2024–25
Do vision-language models pick up the same sound–meaning associations humans do? Code on GitHub ↗
2024–25
Thesis: Cross-Modal Sound Symbolism in Vision-Language Models. Grade 8/10.
2020–23
Minor in Data Science.
WHAT I WORK WITH
WRITE-UPS, NOTEBOOKS AND FINAL MODELS
Predicting daily rainfall from weather features. After feature engineering and K-fold cross-validation, a well-tuned KNN beat heavier ensembles. Finished rank 5 of 2,529.
Regression on house sale prices, using SHAP to see which features drove the predictions. Over several versions I found data leakage in the target, which explained an accuracy jump that had looked suspicious.
Binary classification of loan default risk from credit metrics, debt-to-income ratios and payment history. An ensemble of XGBoost, LightGBM and CatBoost, with SHAP for explainability.
Predicting a track's tempo from raw audio, combining signal-processing features with features based on music theory, then gradient boosting on top.
Interaction features that capture how time of day and weather combine. Training ran in a Docker pipeline tuned with Optuna, so results were reproducible.
Predicting whether a customer responds to a campaign within seven days. A YAML config defined feature sets, models and hyperparameters, making them quick to swap and test.
Estimating how long a listener stays with an episode. Temporal features and a stacked ensemble picked up signal that no single model found.
Recommending fertilizer types per field as a multi-label problem, optimizing directly for MAP@3 rather than plain accuracy.
The classic first competition. I used it to learn feature engineering basics and XGBoost before moving on to messier datasets.
A harder follow-up, and mostly a lesson in overfitting: pushing too hard on validation score hurt generalization until I paid closer attention to feature correlation.
Fraud detection on a heavily imbalanced dataset. The key lesson: apply SMOTE inside each K-fold split, not before it, or the validation scores mislead.
Classifying personality type from a small, noisy dataset. Oversampling and Bayesian hyperparameter optimization made the biggest difference.
Estimating calories burned from workout duration, heart rate and body metrics, engineered into features for gradient boosting.
Predicting exam performance from study habits and demographic features, blending CatBoost and LightGBM instead of relying on a single model.
Stacking CatBoost and XGBoost; weighting their outputs balanced the different ways each model tends to fail.
Multi-class classification of irrigation need from soil composition and weather data.
Predicting the probability that a buyer purchases an electric vehicle, treated as probability estimation rather than yes/no classification.
An agent plays against an opponent in a simulated farm economy (plant, water, harvest, sell) and wins by out-planning the other strategy in real time.
Classifying heart disease risk with a support vector machine, using feature scaling and kernel tuning to get the decision boundary right.
A recurrent neural network on historical market data, to see how much signal a sequence model can find in noisy crypto prices.
Exploratory analysis and classification on a thyroid cancer risk dataset, looking at which clinical features actually separate the classes.
Solving variants of the classic knapsack problem: a break from ML to work through combinatorial optimization by hand.
My own exploration of credit card transaction data, separate from the Kaggle fraud-detection write-up.
Reusable data science helper functions that grew alongside my other projects, and later became a proper utilities library.
Reusable ML components and helpers taken from code I kept rewriting, so common preprocessing and evaluation code doesn't start from scratch each time.
Code for my MSc thesis on cross-modal sound symbolism in vision-language models.
© 2026 Robin Kras · Résumé