ongoing and past research work
formulated memory interference as a failure mode in llm-based multi-agent systems, designing architectural variants for controlled retrieval-scoping.
designed a controlled evaluation framework to quantify output stability of llms under stochastic decoding conditions across 1,500+ tasks.
researching methods to improve diversity, relevance, and ranking stability in recommendation pipelines using llm-driven approaches.
benchmarked cnn and transformer architectures for emotion recognition on fer2013 dataset, analyzing accuracy-efficiency tradeoffs.
improved ndcg/mrr metrics on matrix factorization models through novel regularization and training strategies.