Xiaoyu Yang

Applied Scientist Toronto, Canada

Portrait of Xiaoyu Yang

About

I'm a Machine Learning Engineer at Cresta AI, where I build and productionize LLM systems — fine-tuning and distillation, retrieval and reranking, and agentic Q&A — for quality-critical applications in call-center AI.

Before Cresta, I was an AI Research Scientist at LG Toronto AI Lab, and I completed my Ph.D. in Electrical and Computer Engineering at Queen's University, where my research focused on natural language reasoning with diverse knowledge sources.

My current focus is on post-training methods, agentic retrieval, and rigorous evaluation of LLM-based systems.

Experience

  1. Machine Learning Engineer

    Jul 2023 — Present
    Cresta AI · Toronto, Canada

    Building RAG, agentic retrieval, and in-house LLMs for call-center knowledge assistance.

    • Productionized an in-house query rewriter via LoRA fine-tuning and teacher-model distillation, matching frontier-LLM quality at a fraction of the serving latency and cost.
    • Fine-tuned a lightweight gating classifier that decides when to trigger retrieval, reducing unnecessary LLM calls without hurting answer quality.
    • Improved retrieval relevance through hybrid search, LLM-based document expansion, and in-house reranker fine-tuning.
    • Shipped an agentic retrieval system as the default retrieval layer for production Q&A agents, and built LLM-as-a-Judge evaluators for systematic offline evaluation.
  2. AI Research Scientist

    Oct 2022 — Jun 2023
    LG Toronto AI Lab · Toronto, Canada

    Built an in-house retrieval-augmented language model (Fusion-in-Decoder + distillation) for open-domain QA, with scalable distributed training and sub-500ms retrieval via FAISS + product quantization.

  3. Machine Learning Research Intern

    Jun — Aug 2022
    Microsoft Research · Montréal, Canada

    Researched compositional generalization in semantic parsing, exploring attention regularization and syntax-semantics disentanglement.

  4. Machine Learning Research Intern

    Jun — Aug 2021
    Samsung Research America · Mountain View, USA

    Improved neuro-symbolic visual concept learning with knowledge-aware supervision, increasing model accuracy and training efficiency on visual question answering.

  5. Machine Learning Engineer Intern

    Jun — Aug 2018
    Alibaba DAMO Academy · Hangzhou, China

    Developed a neural information extraction model for insurance policy documents, improving extraction accuracy through attention mechanisms.

Publications

More on Google Scholar →

Education

Honors