Haiyi Xing

Designed and A/B-tested 150+ task-specific prompts, improving answer relevance by 17% and reducing hallucinations by 12%. Curated, de-duplicated, and labeled ~2.3M telecom QA pairs, improving model perplexity from 4.8 → 4.1 (–15%) after retraining. Developed a Python/LangChain evaluation harness now adopted across 3 model-ops teams for regression testing.

Implemented Logistics Regression Model and Gradient Boosting Classifier to design an intention recognition model and could support daily conversation with users. Implemented BERT model, ChatGLM, and Langchain to develop an AI questioning and answering platform to help customers to address issues with their Lenovo devices. Conducted a machine learning presentation to the workers in Lenovo.

Designed an intent recognition engine (Logistic Regression + Gradient Boosting) that raised top 1 F1 from 78 → 92 %, enabling natural small talk and task routing in the support chatbot. Developed a self-service Q&A platform (BERT retrieval + ChatGLM generation + LangChain orchestration) covering ~80% of device issues. Conducted an ML-in-Production workshop for 30+ engineers; training materials became part of the onboarding curriculum.

Selected as one of ~50 global fellows for Alibaba’s flagship immersion program, shadowing teams at Taobao, Tmall, Cainiao, and Alibaba Cloud. Drafted a threat-model brief on personalization vs fraud risk, adopted by the platform-security group. Delivered a talk on “Future of AI in Commerce” to 200+ attendees, ranked as a top 3 speaker at Alibaba’s AI Summit in the program.

Designed an interactive heat-map dashboard for public-opinion monitoring and trend alerts. Migrated backend to Spring Boot + MyBatis, cutting API latency by 68% (380ms → 120ms). Rebuilt frontend in Vue 3 SPA, improving usability and page-load speed.