Malikeh Ehghaghi (/mɒː.liːˈke/)
I'm a first-year PhD student in Computer Science at the University of Toronto, advised by Colin Raffel, and an ML Research Scientist at the Vector Institute. My current work focuses on AI safety and security, grounded in a broader background that spans efficient large-scale pretraining, continual learning, model merging, modular AI, decentralized training, interpretability, and fairness.
Before my PhD, I studied Computer Engineering at the University of Tehran, and later completed my Master's in Applied Computing (MScAC) in Computer Science at the University of Toronto under the supervision of Frank Rudzicz and Jekaterina Novikova.
I spent five years in industry, starting in Toronto in the sensitive domain of AI for mental health at Winterlight Labs and Cambridge Cognition. Building models to detect mental health disorders from speech made one thing clear: "accuracy alone is not enough". In high-stakes settings, systems must be reliable, fair, and deeply interpretable. If a model is not transparent and trustworthy, it is not ready for real-world use.
I later moved to Silicon Valley as a founding NLP engineer at Arcee AI, where we challenged the idea that "bigger is always better." We focused on building efficient, domain-adapted small language models, emphasizing both model-centric approaches (e.g., merging, pruning, distillation) and data-centric approaches (e.g., data mixing, synthetic data generation, filtering, deduplication). That experience shaped a core belief I carry into my research today: the future of AI should not be centralized, opaque, or inaccessible.
After five years in startups, I returned to academia with a clear goal: to bring a lean startup mindset to research — starting with hypotheses, building MVPs, and iterating through build–measure–learn cycles grounded in real-world feedback. I focus on building systems that hold up beyond the lab, guided by a vision of de-risking, democratizing, and decentralizing AI.
I see "AI safety" not as a single problem, but as a stack:
- How we build models? e.g., training efficiency, decentralized learning
- How we align them? e.g., alignment, post-training, and reward modeling
- How we evaluate them? e.g., risk modeling, robustness, and benchmarking
- How they interact with the world? e.g., human interaction, trust, accessibility
Beyond research, I invest in that final layer by building communities for minorities in tech. I serve as the Vice Chair of the ACM-W Toronto Professional Chapter, building a local hub for women in tech in Toronto to share knowledge, connect, and collaborate. I previously co-hosted the Women in AI Research podcast, splotlighting women in AI research. I also founded Techrun, as a shared space for learning, networking, and building across the Persian-speaking tech community in Toronto.