My four-year PhD investigates how large language models are reshaping recommender and content-discovery interfaces. Rather than focussing on the underlying recommendation algorithms, I'm exploring how the interfaces around them can contributes. The aim throughout is to bridge new possibilities with generative AI with genuinely user-centric design. My central research question focuses on composing an interface (its layout, its components, what it foregrounds) to match a reader's actual, evolving information need, and proving that this match matters.

The first phase of my research will map where LLMs can most usefully shape recommender and content interfaces, and just as importantly, where they shouldn't. That work will produce a validated taxonomy of what readers actually need from a complex, evolving story (a quick check-up versus broad exploration, for instance). I love building prototypes for mobile news reading, that compose interface components (timelines, actor maps, comparison views) to match the user need rather than defaulting to one layout for everyone. Some pressing ethical questions are importznt here: the guiding measure is whether the interface helps someone actually explore and understand — not how long it keeps them engaged.

I'm also exploring how to turn tat taxonomy into an evaluation: a controlled comparison of interfaces that match a reader's need against interfaces that don't, alongside a generic relevance-only baseline and the plain article as it exists today. Recommender systems are usually evaluated by whether people click, not by whether the presentation actually served what they came for — this phase is built to answer the second question directly, on mobile, where interface real estate is scarce and the design trade-offs are sharper due to limited screen size.

Positioned at the intersection of two research lines — the long-running academic tradition of adaptive interfaces and the fast-moving, industry-driven push toward generative UI — the project's contribution is the piece the second line is currently missing: a validated theory of which interface structures actually serve which reader needs. The goal: making sure that as recommender and content interfaces get rebuilt around generative AI, they stay accountable to the people using them, not just to engagement metrics.