About
Simon-Shlomo PoilI am Simon-Shlomo Poil. I build systems, products, and organisations around difficult technical problems. My work has moved through neuroscience, biomarker research, and production AI, but the underlying question has remained the same: how do you understand whether a complex system is healthy when its most important properties are not directly visible?
Today I work on production AI in critical infrastructure. Earlier in my career, I worked on neurophysiological biomarkers, EEG-based biomarker analysis and endpoint development for pharmaceutical clinical trials, and the problem of turning validated methods into reproducible workflows and usable systems. Across these settings, the recurring challenge has been how to infer hidden system state from imperfect signals, and how to make that inference robust enough to support real decisions.
Three chapters, one thread#
In academic neuroscience, I studied criticality in neuronal networks, long-range temporal correlations in brain signals, and the use of neurophysiological measurements as biomarkers of cognitive decline. My PhD work focused on how temporal structure can reveal hidden changes in system state, and on how weak individual signals can become more useful when interpreted together in a principled framework. I also built the open-source Neurophysiological Biomarker Toolbox (NBT), which helped make advanced EEG and MEG biomarker methods more accessible and reproducible for other researchers.
In pharma data science, I moved from theory into operational measurement. As co-founder and CTO of NBT Analytics, I worked on EEG biomarker analyses for pharmaceutical clinical trials under quality-assurance constraints. The methods had to be reproducible, the pipelines defensible, and the results usable across different studies, teams, and regulatory contexts. That period made one lesson unavoidable: rigorous methodology only becomes valuable at scale when it can be reused without becoming fragile.
In production AI, I have worked on business-critical machine-learning and data-processing systems. The surface details are different, but the structural problem is familiar. System health is often latent. Drift, feedback loops, changing operating regimes, and human interaction can degrade a system long before average outcomes visibly collapse. Reliable operation depends not only on models, but on observability, monitoring, and the discipline of inferring hidden state without mistaking convenience for truth.
This site collects the research, tools and publications from that work.