Recommendation is where machine learning meets taste at scale. This track is my running investigation into what modern models — especially LLMs — actually add to preference modeling, ranking, and retrieval.
RecSys is where I work through recommender systems as a research problem — experiments, reproductions, and honest evaluations rather than hot takes. The classical stack of collaborative filtering, matrix factorization, and sequential models is always the control group.
I'm most interested in the seam where language models meet ranking: whether an LLM can read behavior as taste rather than IDs, where that helps, and where recommendation's hard constraints — latency, calibration, popularity bias, weak evaluation — push back. The work lives in that tension, not in pretending language models simply replace ranking systems.
Can a model read a session the way a good recommender engineer does — and where does that actually pay off?
Each one takes a question in recommendation and runs it to ground — baselines, experiments, and what the results actually support, written up as I go.