Vania Janet Raya Rios

I've built hybrid retrieval and generation systems at UNAM-IIMAS, where I first-authored a SemEval 2026 paper that earned a Best Paper Award and ranked 4th of 38 systems on IBM's Multi-Turn RAG benchmark. My work combines SPLADE, dense embeddings, Reciprocal Rank Fusion, reranking, and LLM-based query rewriting.

At Mastercard, I engineered a Python and SQL reporting pipeline that reduced cycle time by 80%, while at PagaTodo I optimized vector retrieval to reduce support-pipeline latency by 80%. I'm also developing multilingual clinical retrieval at the University of Zurich and compressed Czech-to-German translation models at MBZUAI, retaining 99.7% of dense-model COMET performance with calibrated 4-bit GPTQ.