Ivan Banny
Deep learning research
Eindhoven, NL
I'm an MSc student in Data Science and AI at TU/e. I work on deep learning architectures, from their mathematical foundations to how they run on hardware. I use PyTorch and JAX, and write C++ and custom CUDA kernels to speed up model execution.
I've published research in AI safety. My background includes IMO selection and ICPC regionals. I'm looking for research roles where I can develop and study new architectures.
I'm also Tech Lead of Serpentine AI, TU/e's AI student team, overseeing all our projects.
Dirtflow
Flow matching for 3D generation and editing.
I independently develop Dirtflow, adapting pretrained TRELLIS models with two-stage flow matching to generate geometry and full Minecraft block states from text and 3D context.
I built the data pipeline for nearly 470,000 builds and the distributed training and evaluation systems. My experiments study learning dynamics and how geometry constrains material generation.
Droomba
World models for autonomous flight.
I'm working on model-based reinforcement learning for drone navigation in unfamiliar environments. The approach learns action-conditioned world models and trains control policies through imagined trajectories.
I'm studying how perception and memory affect generalization to unseen layouts, with transfer from simulation to physical flight as a central problem. The goal is real-time onboard control without a prebuilt map or prescribed route, within the drone's compute and latency limits.
Other work
PCA-assisted Bayesian optimization
Parallel search in high-dimensional spaces using orthogonal subspace sampling. I developed and benchmarked the approach on BBOB functions in 10-40 dimensions.
Motor imagery decoding
Studying how architecture and signal representation affect EEG classification. I compare models under a shared evaluation pipeline and investigate the methodological choices behind published results.
Background
Publications
- 2024An indicator for effectiveness of text-to-image guardrails utilizing the Single-Turn Crescendo Attack ↗
- 2024The Effects of Monetary Policy Shock: Evidence From Systemically Important Economies ↗
Education
-
2025 - presentEindhoven University of
Technology
MSc Data Science & Artificial Intelligence -
2022 - 2025Leiden University
BSc Data Science & Artificial Intelligence -
2024 - 2025TU Delft
Minor in Project Management
Mathematics & algorithms
ICPC regionals in Benelux and Northwestern Europe. Ukrainian IMO selection and 20+ olympiad prizes. Yulia's Dream, the MIT mathematics research program.
Earlier work
Foreground / background separation with a PyTorch CNN at Explore AI @ NYCU, a Google Research supported team program (2023).
Tools
Python, C/C++, CUDA, PyTorch, JAX.
Linux, Slurm,
distributed training.
Docker, AWS, Azure.