Vera Yan

About Vera Yan

Machine learning enthusiast with a strong background in building, improving, and maintaining large-scale machine learning systems.

Experience

1. Machine Learning Engineer

   Meta

   Sep 2021 — Present · Seattle, WA, US

   Smart Glasses Voice AI Assistant Team *- Led post training of Llama-4 powering Meta Ray-Ban and Oakley smart glasses, owning SFT, RLHF, and online RL pipelines end-to-end. Drove 7 training, evaluation, production release cycles, improving task success rate (+16%), capability awareness (+10%), and instruction following (+5%)- Led training and production launch of an LLM-based routing system for Meta smart glasses, orchestrating tool selection and API invocation at runtime. Expanded the system from text-only to multimodal and from English-only to multilingual, while consolidating three router models into a single unified architecture, reducing system complexity and operational overhead- Designed and built an end-to-end data governance platform for LLM post-training, automating privacy compliance, data retention enforcement, schema validation, and text-to-speech preprocessing. Onboarded 100+ datasets across multiple teams and eliminated significant manual engineering effort, unblocking scalable and compliant model iteration.* Bot Detection Machine Learning Team *- Identified and addressed core bottlenecks in large-scale ML systems for scraping bot detection. Led a team of 3 engineers to improve label quality, introduce new client-side and behavioral signals, and migrate models from gradient-boosted trees to sequence models (LSTM), increasing true positives by ~25% at constant false-positive rates- Led a strategic shift from per-endpoint bot detection to clustering semantically similar API endpoints using sequence embeddings (word2vec). Reduced model count from tens of thousands to fewer than 10, dramatically improving system scalability and long-term maintainability across an evolving adversarial space- Led 2 engineers to design and implement a production ML evaluation and monitoring framework from scratch, providing end-to-end visibility into model performance and reducing team on-call burden by ~50%.

Education

• Duke University

  Master of Science (MS), Statistical Science

  2016 — 2018

• Nankai University

  Bachelor of Arts (B.A.), Finance, General

  2012 — 2016