Jun 2024 – Dec 2024 · 6 mos · Peking, China · On-site

• Led the intelligent assistant's auto-ordering project within ACT (transformer for ACTions), utilizing Qwen-VL and an XPath-based user action path decision tree. Improved execution plan generation and replication accuracy to 85%, with plans to scale the automation method to millions of Alipay mini-programs.
• Designed the GraphRAG recall and embedding algorithms for Alipay's intelligent Q&A system, and explored the general application of RAG techniques across various data sources.

Jun 2023 – May 2024 · 11 mos · Peking, China · On-site

• For the task of generating high-quality comments on short videos, designed a pipeline for training data filtering, multimodal fine-tuning, and model evaluation based on the mPlug framework. Achieved an AUC of 0.79 in the perplexity ranking of human-labeled comments, and deployed it to support real business use.
• Explored the multimodal capabilities of large language models (LLMs) at the output level, designing the corresponding architecture and training approach.

Dec 2020 – Mar 2021 · 3 mos · Hangzhou, Zhejiang, China

• • As a researcher, independently designed a multidimensional feature fusion approach for millimeter-wave technology. A research paper on multi-person pose reconstruction based on Texas Instruments’ commercial millimeter-wave radar was published in MobiSys 2022 (CCF-B).

May 2020 – Jun 2021 · 1 yr 1 mo · 中国 上海市 闵行区 · On-site

• Multi-modal sensing and modeling.