ML Engineer · Vienna, Austria
I build ML systems end-to-end — from data pipelines and model training to containerised deployment. Four years of research at IST Austria and IIT Madras, two publications in Nature Scientific Reports.
Selected works
Built and deployed across neuroscience, deep learning, and reinforcement learning. Each one live and interactive.
Deep autoencoder learning compact spatial representations from 250k 3D trajectory samples. Task-relevant structure emerges without explicit supervision — validated across two experimental paradigms, published in Nature Scientific Reports.
Multi-task CNN–GNN trained on 1M+ synthetic trajectory points via a custom Unity3D simulation pipeline. 7-config ablation on prediction targets; continual learning with EWC regularisation across 7 sequential sessions.
End-to-end RL pipeline from scratch — custom 49-neuron state encoder at >90% position reconstruction accuracy, TD-trained value network, hill-climbing inference. No explicit path planning.
I started in aerospace, ended up in computational neuroscience, and spent four years turning research questions about how the brain represents space into production ML systems. The gap between what is published and what is deployed is where I am most comfortable.
When I replaced a manual workflow at IST Austria that took 3–4 months per session with a fully scripted pipeline running in 30 minutes, the most useful realisation was not about the pipeline itself — it was about how much of what we called research infrastructure was just engineering debt nobody had named yet.