Projects

Explored how model quantization impacts machine translation (MT) performance across varying sentence difficulty levels. Using quantized variants (8-, 4-, and 2-bit) of TowerMistral and Gemma-3, evaluated across four languages, we examine how reduced precision influences both semantic and surface-level translation quality.

Investigated whether learned 3D geometry from RGB images can replace depth sensors in few-shot robotic manipulation. Integrated MASt3R, a state-of-the-art 3D reconstruction method, into the RVT-2 robotic vision pipeline to evaluate performance without RGB-D sensors. Demonstrated that learned geometry can enable robust manipulation under challenging visual conditions. Possibly offering a more flexible, cost-effective alternative to traditional calibrated setups.

AI for Dummies is an educational initiative I started to explain AI concepts to beginners. Sparked by growing interest and common misconceptions observed during my work at HvA, the project aims to deliver intuitive lessons and providing backup to workshops on topics like machine learning, probability, and generative AI.

Investigated the emergence of compositional language in referential games between artificial agents, using Graph Neural Networks (GNNs) to model structured input. Explored how agents develop communication strategies to describe family relations, analyzing the impact of different training setups on language expressiveness and efficiency.