projects
SNSF Improving Public Discourse
Program. Improving Public Discourse — Swiss National Science Foundation grant hosted at the University of Zürich Digital Democracy Lab (PI: Fabrizio Gilardi).
My role. Technical lead. I design and maintain the data collection and classification infrastructure across the program's studies.
Current study. Audit study of TikTok content moderation. Research assistants systematically review and flag content violating community guidelines; the platform's enforcement response is recorded and analyzed. Study launched March 2026 and is actively running.
Methods. Audit methodology, two-stage NLP classification pipeline (multilingual BERT + 9B-parameter LLM), Postgres + Redis infrastructure, Ansible deployment, Flask RA/admin UI.
Collaborators. Karsten Donnay (PI), Fabrizio Gilardi (program PI), Daria (classification pipeline), UZH DigDem Lab RAs.
project site ↗
Dissertation
Title. "Governance, Competition, & Extremism: How the Structure of Social Media Platforms Radicalizes Communities"
Core argument. Platform governance structures create competitive dynamics between communities that systematically drive radicalization. A comparative study of digital architectures and how they shape collective behavior.
Methods. Multi-platform analysis (Reddit, Twitter, Telegram, TikTok, YouTube), transformer models, network analysis, digital ethnography.
Advised by. Jennifer Larson (co-chair), Cassy Dorff (co-chair), Emily Ritter, Anita Gohdes (Hertie School).
job-market-paper.pdf ↗
VIRENA
Title. Indirect Effects of Content Moderation Errors: A Chatroom Experiment with AI Agents
Description. VIRENA is an experimental platform simulating WhatsApp-style social media chatrooms. Prolific participants interact in real-time with LLM-powered bot personas under randomized moderation conditions — over-moderation, under-moderation, and regular moderation. The experiment tests how moderation errors shape conversation dynamics and perceptions of online spaces.
Status. 800-participant study live. Platform rebuild (VIRENA) underway for future experiments.
Stack. Svelte.js (frontend), PocketBase + Go (backend), Ollama / OpenAI (bot personas), Prolific (participant recruitment).
Collaborators. Karsten Donnay (PI), Giuliano Formisano, Emma Hoes, Jonathan Klues, Adiba Mahbub Proma, Fabrizio Gilardi.
Platform ABM
Title. Voting with Your Feet Online: Platform Competition, Extremism, and the Limits of Tiebout Sorting
Core argument. Online platforms function as a system, not independent sites. Applying Tiebout's "voting with your feet" model to digital platforms, the paper shows that extremist communities act as utility vampires — degrading mainstream experiences and forcing displacement, raiding cycles, and ideological enclaves. The diversification premium (more platforms = better outcomes) grows with extremist threat intensity, but governance type determines who benefits.
Methods. Agent-based model, three-way ANOVA, bootstrap CIs, sensitivity analysis (OAT + Sobol indices). Python simulation; R/Python visualization.
Status. Working paper — manuscript drafted, sensitivity analysis complete, figures integrated. Preparing for submission.
Related. Ideological fragmentation of the social media ecosystem (PNAS Nexus)
