Projects & Industry Experience
This page collects the applied projects and workstreams highlighted in my research CV.
Agent Harness 驱动的多模态质量诊断与自进化
Algorithm Engineer · Multimodal, Taobao Tmall Group / China E-commerce Business Group.
Built trajectory compression, evidence routing, and Attribution/Critic/Counterfactual Grader rubrics for 300+ node Agent chains. Batch verification reached nearly 95% high-value case reduction and moved end-to-end diagnosis from tens of minutes to the minute scale. Also developed a severity-aware SOP alignment and fine-tuning loop with SFT samples and graded rewards.
数据基础设施与安全事业部 · 算法开发
Built a rule/Agent dual-path cleaning system for network-security assets, then designed fragment-indexing and semantic/fine-grained matching algorithms for asset-vulnerability retrieval. The work improved search logic, recall quality, coverage, and pipeline maintainability.
华为通用向量检索合作项目
Core contributor to hybrid vector retrieval and the STABLE project. The collaboration received the Huawei Outstanding Technical Collaboration Award and is connected to the industrialization of high-throughput heterogeneous vector search.
Research Skills
- Agent and feedback loops: Agent Harness, trajectory/trace compression, evidence routing, Grader/Rubric, Critic, self-play, SFT/DPO/RLVR, Skill/Memory.
- Multimodal and retrieval: MLLM task modeling, composed image retrieval, vision-language representation, ANN, hybrid metrics, heterogeneous indexes, dynamic routing.
- Trustworthy AI: AIGC generation and detection, artifact elimination, AI watermarking, safety evaluation, data cleaning, index construction, and batch evaluation.