Denys Herasymuk

I am a PhD student at the NYU Center for Data Science, working at the Center for Responsible AI with Prof. Julia Stoyanovich. My research is on AI safety for temporal intelligence: whether models reasoning over dynamic systems deserve trust, what their confidence is worth, and how they hold up as conditions change. I study time-series language models in high-stakes settings like the ICU, where failure can mean a missed deterioration or harm to patient health.

I came to these questions from the responsible-AI side. As a research fellow at NYU before starting my PhD, I examined how machine learning models behave beyond average accuracy: how stable, fair, and interpretable their predictions remain across different populations and data distribution shifts. What stayed with me is that the failures that matter most tend to surface after deployment rather than on benchmarks.

At Seattle Children's Research Institute I moved to physiological time series, working alongside clinicians on models built from real hospital biosignals. That is where I learned what safety means in practice: benchmark success is not readiness, demographic groups underrepresented in training data are most affected, models stay overconfident and miscalibrated, and they still lack a dynamic patient representation — reading one recording at a time rather than a longitudinal trajectory.

I also completed a postgraduate research program at Imperial College London, where I worked on privacy and robustness in decentralized learning. Earlier, I studied computer science at the Ukrainian Catholic University, and alongside research I spent several years building large-scale data infrastructure in industry — which is where my attention to what happens after a model ships comes from.