Curriculum Vitae
Summary
Master's student in the Machine Learning Department at Carnegie Mellon University focused on reliable and interpretable AI, multi-agent training paradigms, and activation-based model analysis.
Education
Experience
Advised by Shi Feng. Developed persona-training methods which emulate midtraining for better model organisms and constitutions. Engineered evaluations across 100+ training runs on synthetic-document pipelines. Evaluated how tastefully frontier models can test ideas in safety research.
Develop multi-agent reinforcement learning environments to improve legibility of reasoning traces. Design user studies and adversarial setups to measure influence of legibility on human and multi-agent interactions.
Engineered activation profiles of language model evaluators to suppress self-preferential bias, applied to LLM routers (ICML 2026, MechInterp @ NeurIPS 2025). Winner of Mechanistic Router Interpretability Hackathon (40 submissions).
Designed and trained large-scale graph-text models; evaluated adversarial robustness of LLMs; built historical correspondence networks and dashboards.
Automated ETL pipelines and implemented synthetic-data imputation for compromised client data.
Deployed decision-management systems and structured-prediction models on meeting transcripts.
Built Amazon sales data visualizations and automated analytics with GCP BigQuery and BERT-based embedding search.
Extended NER to long-tail ethnics cuisines using Bon Appetit data.
Publications
Conference Papers
Dani Roytburg, Matthew Bozoukov, Matthew Nguyen, Jou Barzdukas, Mackenzie Puig-Hall, and Narmeen Oozeer. Are LLM Evaluators Really Narcissists? Sanity Checking Self-Preference Evaluations. Forty-Third International Conference on Machine Learning, 2026.
Dani Roytburg, Shreya Sridhar, and Daphne Ippolito. Measuring Weak-to-Strong Legibility of Reasoning Models. Advances in Neural Information Processing Systems (NeurIPS), 2026.
Dani Roytburg and Daphne Ippolito. Disentangling Models from Personas in Heterogeneous LLM Simulations. Second Workshop on Social Simulation with LLMs: Fidelity in Applications (COLM 2026, Spotlight), 2026.
Dani Roytburg* and Beck Miller*. Mind the Gap: Pathways Towards Unifying AI Safety and Ethics Research. Proceedings of the International Association for Safe and Ethical AI, 2026.
Dani Roytburg*, Matthew Bozoukov*, Hongyu Fu, Matthew Nguyen*, Jou Barzdukas*, and Narmeen Fatimah Oozeer. Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLM Evaluators. Mechanistic Interpretability Workshop at NeurIPS 2025, 2025.
Dani Roytburg*, Deborah Olorunisola*, Sandeep Soni, and Lauren Klein. Words and Action: Modeling Linguistic Leadership in # BlackLivesMatter Communities. Proceedings of the International AAAI Conference on Web and Social Media, 2025.
Awards & Recognition
- Best Poster, LTI Student Research Symposium, 2026
- Emory University Dean's List, 2022, 2024, 2025
- Winner, Martian Research Mechanistic Interpretability Hackathon, 2025
Skills
Python · Java · R · JavaScript · Typescript · SQL · PyTorch · JAX · HuggingFace Transformers · scikit-learn · spaCy · networkx · D3.js · React.js · GCP · Docker · MySQL