Researching subliminal learning and data poisoning in model distillation.
Earlier work: language models for code, student–AI interaction research, and scientific simulation.
Open to research engineering and applied ML roles from autumn 2026.
I'm Tim, a Finnish–Ukrainian MSc student at UCL. My thesis looks at subliminal learning and data poisoning during model distillation — specifically, what can pass from a teacher to a student when the training data looks innocuous.
I tend to gravitate towards work that sits awkwardly between research and implementation. At TNO-ESI, that meant evaluating open and proprietary LLMs for industrial code graphs. At VU Amsterdam, it meant putting an AI support agent into a large course Discord and building simulation tools with climate researchers.
I also care a lot about education. That is partly professional and partly personal, shaped by finishing school during COVID and seeing education disrupted by the war in Ukraine.
A selection of research and software projects.
Investigating whether hidden behaviours can transfer from teacher to student models through sequence-level distillation, and what that reveals about realistic anti-distillation attacks and their defences.
Built a Text-to-Cypher evaluation dataset and compared open and proprietary LLMs. The key result: on a specialised schema with limited data, base-model capability mattered more than small-scale fine-tuning.
Contributed to CuReMix for COMP0258: Open-Endedness and General Intelligence. The project treats intrinsic and extrinsic reward weighting as an automatic curriculum; across sparse-reward Craftax tasks, it matched reward mixtures tuned separately for each task without task-specific tuning.
Turned climate and salinity research into an interactive Godot simulation for policymakers, farmers, and students. The project won the SALAD & Trustfarm hackathon and was shown at FOSC 2024 in Marrakesh.
TNO-ESI
Evaluated LLMs as natural-language interfaces for high-complexity code-analysis tooling.
Vrije Universiteit Amsterdam
Supported a longitudinal study of how students use AI in academic settings.
Vrije Universiteit Amsterdam
Translated climate and salinity research into an interactive VR simulation.
Vrije Universiteit Amsterdam
Taught Python, Intelligent Systems, and Applied Programming for AI.
Finnish Defence Forces · Navy
Maintained simulation software and hardware for military training environments.
University College London
Vrije Universiteit Amsterdam