MSc Computer Science, LIACS, Leiden University · Netherlands · LinkedIn

Over a decade in industry, starting on mainframe systems and moving through data engineering, analytics and machine learning, most recently in senior and lead roles. I returned to full-time study to go deeper on the modelling side. My thesis is on natural language inference and vision-language understanding, and I also work on deep learning and reinforcement learning.

Open to Data Scientist, Machine Learning Engineer, AI Engineer and Data Engineer roles in the Netherlands.

This account holds my academic and personal work; professional work is under employer NDA and is covered on my CV and LinkedIn. Every repository below leads with what the experiments actually showed — including the results that went against what the method predicted — and ships the full report alongside the code.

• atomic-visual-entailment — master's thesis: decomposing hypotheses into atomic facts for interpretable visual entailment
• cross-encoder-reranking — cross-encoder re-rankers on MS MARCO with rank fusion and LLM query expansion
• dutch-manifesto-multilabel-classification — multi-label thematic classification with RobBERT-2023 and a long-tail ablation
• text-mining-assignments — text categorization on 20 Newsgroups and BERT-based NER on archaeological reports

• tell-the-time-cnn — five ways of representing clock time for a CNN; the output representation moves error from chance to 8.5 minutes
• generative-models-and-seq2seq-rnns — CAE, VAE and GAN on anime faces, plus encoder–decoder LSTMs for arithmetic
• digit-classifiers-and-xor-backprop — distance-based classifiers, a multi-class perceptron and hand-derived backpropagation

Three studies on CartPole-v1, from value-based methods through policy gradients to PPO.

• dqn-variants-cartpole — target networks and experience replay, separately and together, with a hyperparameter ablation
• policy-gradient-methods-cartpole — REINFORCE, Actor-Critic and A2C against a DQN benchmark
• ppo-cartpole — PPO with a clipped objective, GAE and advantage normalisation

• azure-video-watermarking — five queue-coordinated microservices watermarking video in parallel on Azure
• container-autoscaling — a custom load balancer and scaling controller tested against a latency SLA

• amazon-product-recommender — hybrid recommender under extreme sparsity, where 93% of users have a single interaction
• landmark-shortest-path-estimation — approximate shortest-path distances in large graphs through landmark selection
• surrogate-hyperparameter-optimization — SMBO, Successive Halving and Random Search against a Random Forest surrogate

• Computational-Creativity-Cookie-Generator — a genetic algorithm evolving novel cookie recipes
• bad-poets-society — a poem generator with similarity filtering to stop it copying its own corpus
• Flappy-Bird-Co-Creative — a co-creative level design game