Senior Algorithm Research Expert at the Innovation Division of Ping An Technology
Former Head of Algorithms for Ant Group’s Zhixiaobao and Maxiaocai, where he led the development of the core algorithm system for an intelligent financial advisor serving hundreds of millions of users. He was also a founding core member and chief architect of the Baidu App, deeply involved in building Baidu’s mobile ecosystem from the ground up to hundreds of millions of users. With over 15 years of hands-on R&D and architecture experience in AI, he has led teams to win five international AI competition championships and has published more than 20 top-conference papers and patents. During his time at Intel, the SE-OS technology solution and source code he designed and implemented were procured by the U.S. National Security Agency. He is currently focused on algorithm and agent innovation for AI-native applications in healthcare and finance, aiming to deeply integrate cutting-edge large model technologies with specialized domains to build highly reliable and deployable AI-native products that drive industrial intelligence transformation.
Topic
Breaking the “Impossible Triangle” of Large Model Deployment: Ping An Technology’s Multi-Agent Practice in Complex Medical Scenarios
Professional healthcare is the ultimate test for large models. In clinical settings, where uncompromising accuracy is required and knowledge is highly dependent on expert experience, general-purpose large models often fall short—the “hallucination rate” remains high, clinical knowledge is largely implicit, and diagnostic pathways are extremely fragmented. Together, these factors form the “impossible triangle” that constrains AI adoption in medicine. This talk breaks away from the traditional AI medical approach dominated by retrieval-augmented generation (RAG) and explores how to rethink AI for healthcare in the LLM era. The presentation will introduce, for the first time, Ping An Technology’s three core algorithmic engines developed during the creation of its critical illness MDT diagnostic product and the results achieved: Expert Clinical Knowledge Internalization Engine Reveals how implicit expert knowledge and clinical experience can be converted into model-learnable reasoning paths, bridging the gap from “knowledge retrieval” to “experience-driven inference.” Multi-Dimensional Reasoning Evolution and Hallucination Control Engine Addressing the complexity of medical reasoning and the need for low hallucination, a three-stage progressive training strategy was designed: static chain-of-thought alignment, dynamic chain-of-thought reinforcement, and high-order reasoning leaps. This enables models to perform incremental long-chain reasoning and interdisciplinary inference in complex critical illness scenarios, while fundamentally suppressing hallucinations to ensure accurate and reliable reasoning. Ping An Critical Illness Multi-Agent Collaboration Engine Explains the evolution from single-agent to multi-agent collaborative architecture. By entering deep research mode, the system automatically processes massive literature, conducts cross-study evidence comparison, contradiction analysis, and evidence-grade integration, ultimately generating dynamic, personalized diagnostic and treatment plans that help physicians make cutting-edge, evidence-based decisions beyond standard guidelines.
