BSc Data Science & AI @ Maastricht University
Junior Data Scientist · ML Researcher · Builder
I'm a Data Science & AI graduate at Maastricht University with hands-on experience spanning financial analytics, NLP pipelines, and medical ML research. I enjoy building things that work - from CNNs from scratch to production-grade RAG systems and real-time ML services.
- 🔬 Thesis @ MUMC+ - Predicting Postoperative Delirium from intraoperative EEG signals
- 💼 Junior Data Scientist @ LARKinfolab - GraphRAG architecture with ReAct agents
- 🏦 Data Science Intern @ Luminor Group - Large-scale EDA (10M+ rows), built NLP extraction pipeline for 200K+ records, customer segmentation that reduced SMS costs
| Project | Description | Stack |
|---|---|---|
| Fraud Detection | Real-time credit card fraud scoring on 13M+ transactions - XGBoost behind a FastAPI endpoint that reconstructs features from Redis in milliseconds, MLflow model registry with SHAP/feature-importance logging, hourly Evidently drift checks, Prometheus + Grafana dashboards, and CI/CD that ships to production on a git tag | Python, FastAPI, Redis, PostgreSQL, MLflow, Docker, Evidently, Prometheus/Grafana |
| Project | Description | Stack |
|---|---|---|
| CNN from Scratch | Convolutional neural network built with NumPy only - no PyTorch/TensorFlow. Trains on MNIST with custom conv, ReLU, softmax, backprop | Python, NumPy |
| Multilayer Perceptron | Handwritten digit recognition from a single image using a neural network built from scratch | Python, NumPy |
| Simple Language Model | Feed-forward neural language model in PyTorch. Fixed context window of 5 BERT token IDs predicting the next token on tinyshakespeare | Python, PyTorch |
| Competition | Score | Approach |
|---|---|---|
| Stellar Classification (S6E6) | 0.954 balanced accuracy | XGBoost + Optuna + sklearn Pipeline + Error Analysis |
| Project | Description | Stack |
|---|---|---|
| NLP Labs | Labs covering Tokenization (BPE), Information Retrieval (TF-IDF), BERT, Prompt Engineering | Python, HuggingFace |
| Comparative Study NMT | Analysis of NMT architectures: Seq2Seq LSTM vs MarianMT vs M2M100 - accuracy, noise robustness, adaptability | Python, PyTorch |
| Project | Description | Stack |
|---|---|---|
| Yinsh game AI | Yinsh board game with two AI bots: Alpha-Beta pruning and Deep Q-Network | Java, JavaFX |
ML/AI
XGBoost LightGBM scikit-learn PyTorch HuggingFace LangChain optuna SHAP
LLM & RAG
ReAct Agents GraphRAG Neo4j Pydantic local models
MLOps / Infra
MLflow Docker Redis PostgreSQL Prometheus Grafana Evidently GitHub Actions
Data
pandas numpy SQL matplotlib seaborn plotly
Engineering
Python Java FastAPI Docker Git Azure
- 📐 Deepening my statistics foundations
- 🕸️ Building out GraphRAG @ LARKinfolab
- 📝 Writing about ML on Medium
"Train hard, generalize harder"