Thomas Vos
MSc AI student at the University of Amsterdam, working on interpretability, sparse attention and generative modelling. I study how models find their way down loss peaks, and spend my holidays going up real ones.
About
I'm an MSc Artificial Intelligence student at the University of Amsterdam. My research spans interpretability, sparse attention and emergent communication, and lately I've been drawn to generative modelling and flow matching.
Alongside my studies I've spent three years as a part-time machine learning engineer, shipping a production recommender and running long A/B tests. Away from the screen I'm usually on an alpine route, or making music.
A production recommender on AWS, trained on data from more than 3,000 restaurants. I ran a 7-month Bayesian A/B test of a price-capped candidate filter to measure what it actually changed.
+15%
conversion, 3.9% → 4.5%
+18%
add-to-cart
+6.6%
revenue per recommendation
An interpretable churn-risk model built on noisy, heavily imbalanced data. The customer success team uses it to decide which clients to reach out to first, and can see why each one was flagged.
θ₁
Reproduced Matryoshka sparse autoencoders on CLIP and found insufficient evidence for the claimed hierarchy. Removed always-on neurons with a targeted L1 penalty, and introduced a concept saliency method that localises SAE features at pixel level.
research projects & thesis
Course project · 2026
University of Amsterdam · Interpretability & Explainability
A benchmark for whether audio-vision models link speech to the right on-screen speaker. ImageBind matched voices to faces 76% of the time zero-shot (chance is 50%), and probing showed that high latent alignment does not mean better grounding.
Course project · 2026
University of Amsterdam · Deep Learning 2
Adapted static and dynamic sparse attention from LLMs to vision transformers, with custom GPU kernels. Matched the ABMIL baseline on CAMELYON16 (93.8% accuracy), trained up to 47% faster than dense attention and scaled to millions of tokens.
BSc thesis · 2025
Supervised by Dr. R. G. Alhama (ILLC)
Emergent communication in referential games. Pairing a trained sender with untrained receivers gave higher, more stable accuracy and a more compositional language than the controls.
Every step taken with momentum, and a learning rate that was occasionally too high.
Thesis on emergent communication; projects on real-time EEG classification for drone control and an agent-based simulation of Covid-19 spread in supermarkets.
A “busyness” regression model that anticipates errors in existing forecasts, cutting mean departure-time bias from 6.3 days to under 4 hours.
Deployed a production recommender on AWS trained on data from 3,000+ restaurants. A 7-month Bayesian A/B test lifted conversion by 15%, and an interpretable churn model helps the customer success team prioritise.
Interpretability, sparse attention, and more and more generative modelling and flow matching.