Ant Group
July 2018 - Present
LLM Document Intelligence
March 2025 - Present
Business Lead
Independently incubated and led the company's first commercial project for large model document intelligence targeting overseas markets from scratch (0 to 1). Fully responsible for the entire lifecycle from market research, customer requirement discovery, MVP prototype design and validation, to project initiation, R&D management, and final delivery and implementation.
- Multi-modal Document Understanding: Built an understanding engine based on fine-tuning multi-modal LLMs, achieving high-precision structured information extraction from complex layouts like forms and charts.
- Innovative Document Forgery Detection: Pioneered and implemented a document forgery detection solution based on code generation and reinforcement learning. This solution enables the model to autonomously generate and execute image processing and computational code to deeply analyze the micro-features of a document, identifying tampering and forgery traces that are difficult for traditional techniques to detect.
- Agentic AI for Intelligent Document Insights: Utilized a combination of Prompt Engineering and Context Engineering to enable an AI Agent to autonomously understand complex user instructions and perform advanced tasks such as document Q&A, information comparison, content summarization, and risk insight.
ZOLOZ Digital Technology Overseas Commercialization
July 2020 - Present
Algorithm Engineering Lead
ZOLOZ, an independent global trusted identity platform incubated by Ant Group (recommended in Gartner® reports), achieves annual revenue in the hundreds of millions. Focuses on overseas ToB security scenarios like eKYC, risk control, marketing, etc., with core algorithms covering face recognition, document recognition (OCR), AI Agent, and recommendation systems.
- Container Image Acceleration Optimization: Led the refactoring based on the
OCIV2 image format, innovatively adopting file-system-level chunk storage, achieving over 80% reduction in storage resource usage. Designed an on-demand loading mechanism and P2P cache acceleration, enabling efficient management of large model files and second-level container startup, significantly improving cluster resource utilization and service response efficiency. Key contributions open-sourced, recommended by CNCF, and presented at the 2023 OpenAnolis OS Conference.
- Inference Performance Acceleration: Algorithm Orchestration Layer: Utilized
DAG for multi-model parallelism and pre/post-processing optimization. Architecture Layer: Employed message queues to build integrated online/offline inference services. Model Layer: Implemented operator fusion and low-precision quantization, adapting to various inference frameworks and heterogeneous computing resources (Intel/ARM/AMD) for cost optimization. Hardware Level: Collaborated closely with Intel, ARM, and AMD teams to deeply exploit hardware potential. Achievements shared at the 2024 OpenAnolis OS Conference and 2024 Apsara Conference, receiving joint PR coverage from Alibaba Cloud and Intel.
- Large Model Inference System Design & Implementation: Built multi-level operator abstractions for heterogeneous hardware (
GPU/CPU). Optimized inference resources with a cloud-native Parameter-Data separation architecture, enabling efficient KVCache transmission and elastic scaling in multi-cloud environments. Achieved unified resource management and operation for Kubernetes clusters across multiple overseas clouds and sites.
IIFAADevice Biometrics
July 2018 - July 2020
Backend Engineer
IIFAA (Internet Identity Fast Authentication Alliance) focuses on developing and promoting standards for identity authentication and biometric technology, covering over 1.6 billion devices and serving over 1 billion users.
- Responsible for the backend development and maintenance of device-side biometric applications (fingerprint, face, iris, etc.) within the Ant and Alibaba ecosystems, applied in core scenarios like payments, logins, and authorizations, ensuring stability during major promotions like Double 11 and Double 12.
- Established deep collaborations with major domestic mobile phone manufacturers, led the formulation of the
IIFAA protocol standard, supporting over 800 device models. Actively participated in cooperation with the international FIDO Alliance and led the drafting of the international standard for Secure Camera, effectively defending against device injection attacks through hardware-level security mechanisms.
- Collaborated with ISV partners to drive commercialization, successfully serving over
100 APP clients adopting IIFAA standards.
