AI safety research.
Applied ML since 2018.

I’m Sofiia, an AI safety researcher working on LLM evaluation, model incentives, and revealed preferences.

I have worked in ML/NLP product R&D since 2018. That has meant building and evaluating production systems in domains where the data is messy and reliability matters: social media at scale, clinical reporting, urban analytics, transport, and low-resource language research.

At SPAR, I lead the Quora workstream for AI Revealed Preferences. We use forced-choice experiments to surface preferences that are difficult to see in direct model reports. The strongest pattern across 20 tested models was a consistent aversion to “uncomfortable-truth” tasks, a form of covert sycophancy relevant to deception and scalable oversight.

I am also a MARS V fellow at the Cambridge AI Safety Hub, extending the revealed-preferences and incentive-sensitivity research toward tighter experimental isolation and a publishable output.

Since August 2026, I have also been a TALOS Fellow. My work there focuses on structured transparency for safety evaluations and on how AI becomes embedded in social decision-making processes.

Before AI safety research, I built production NLP and LLM systems, clinical-reporting pipelines, urban computer-vision models, and city-scale mobility analyses. I am based in Istanbul.