CV

教育经历

Education

工作经历

安徽华米信息科技有限公司 - AI大模型专家 (2024.11 - 至今)

• 主导Amazfit智能手表语音助手Flow的AI云端开发，实现基于大模型的智能交互功能
• 负责产品功能定义与技术方案设计，协调跨部门团队完成需求落地
• 构建大模型评估体系，优化模型性能指标30%以上
• 开发自动化测试框架，提升模型迭代效率40%

北京原子回声智能科技有限公司 - 大模型负责人＆联合创始人 (2022.04 - 至今)

• 主导研发Atom-1B/7B/13B系列大语言模型，覆盖轻量到通用场景需求
• 建立200TB+中文语料清洗体系，完成多轮数据去重与敏感信息过滤
• 构建从数据清洗到模型部署的全流程技术体系，提升团队研发效率30%+
• 作为Llama中文社区联合发起人(GitHub星标14k+)，贡献核心开源代码与中文预训练方案

北京循环智能科技有限公司 - 算法研究员 (2020.03 - 2022.04)

• 研发自然语言处理算法，优化文本分类与实体识别模型准确率15%
• 设计大模型训练框架，支持千亿参数模型的分布式训练
• 主导多个企业级NLP项目落地，包括智能客服、内容审核等场景

Work Experience

Huami Technology - AI Large Model Expert (2024.11 - Present)

• Led AI cloud development for Amazfit smart watch voice assistant Flow, implementing LLM-based intelligent interaction
• Designed product features and technical solutions, coordinating cross-functional teams for implementation
• Built model evaluation system, improving key performance metrics by 30%
• Developed automated testing framework, increasing model iteration efficiency by 40%

Atom Echo - Large Model Lead & Co-founder (2022.04 - Present)

• Spearheaded development of Atom-1B/7B/13B LLM series for diverse application scenarios
• Established 200TB+ Chinese corpus cleaning pipeline with deduplication and sensitive data filtering
• Built end-to-end technical framework from data to deployment, boosting team efficiency by 30%
• As co-founder of Llama Chinese community (14k+ GitHub stars), contributed core code and Chinese pretraining solutions

Recurrent AI - Algorithm Researcher (2020.03 - 2022.04)

• Developed NLP algorithms, improving text classification and NER accuracy by 15%
• Designed large model training framework supporting distributed training of 100B+ parameter models
• Led implementation of enterprise NLP projects including smart customer service and content moderation

项目经历

Atom系列大语言模型

• 主导研发Atom-1B/7B/13B等模型，覆盖轻量到通用场景需求
• 建立200TB+中文语料清洗体系，完成多轮数据去重、敏感信息过滤
• 构建推理服务引擎，支持32K上下文长度，实现高效部署
• 模型HuggingFace下载量超20万次，落地知识库报告生成、内容生成等多场景

AskOnce个性化知识库产品

• 基于Atom大模型的企业级知识管理平台
• 支持PDF/Word/Excel等多格式文档解析，构建动态知识库索引
• 开发自然语言交互界面，响应时效压缩至秒级
• 建立标准化知识萃取流程，员工问题解决效率提升60%+

宝马企业内部资料问答系统

• 实现企业文档自动化归纳分类与结构化处理
• 开发定制化检索增强生成(RAG)架构，答案准确率达行业领先水平
• 支持与企业OA/CRM系统集成，形成内部知识流转闭环

涉密项目内部资料问答系统

• 实现日均百万级多语言涉密文档自动化解析
• 集成涉密数据分级管控机制，保障敏感信息处理合规
• 建立数据溯源追踪系统，完整记录文档处理全链路日志

通用数据治理算法能力平台

• 支持文本/音频双模态数据治理全流程
• 集成20+核心算法能力：文本分类/实体抽取/翻译/合规检测等
• 覆盖金融、法律、医疗等多领域数据治理需求

分类驱动电销实时话术推荐系统

• 预处理百万级历史对话文本，提取500+高频应答话术
• 实现毫秒级话术推荐响应，减少人工决策时间30%
• 落地电销外呼及客服IM场景，提升沟通效率与标准化水平

Projects

Atom Series Large Language Models

• Led development of Atom-1B/7B/13B models covering lightweight to general scenarios
• Established 200TB+ Chinese corpus cleaning system with deduplication and sensitive info filtering
• Built inference service engine supporting 32K context length for efficient deployment
• Models downloaded over 200,000 times on HuggingFace, applied in report generation and content creation

AskOnce Personalized Knowledge Base

• Enterprise knowledge management platform based on Atom LLM
• Supports multi-format document parsing (PDF/Word/Excel) with dynamic indexing
• Developed natural language interface with second-level response time
• Established standardized knowledge extraction process, improving efficiency by 60%+

BMW Internal Q&A System

• Automated classification and structuring of corporate documents
• Developed custom RAG architecture with industry-leading accuracy
• Integrated with OA/CRM systems to form knowledge circulation loop

Confidential Project Q&A System

• Processed millions of multilingual confidential documents daily
• Implemented multi-level data security controls for sensitive information
• Established full-chain audit logging for document processing

General Data Governance Platform

• End-to-end governance for text/audio multimodal data
• Integrated 20+ core algorithms: text classification, entity extraction, translation, etc.
• Served finance, legal, healthcare and other industries

Sales Conversation Recommendation System

• Processed millions of historical sales dialogues to extract 500+ responses
• Achieved millisecond-level recommendation response, reducing decision time by 30%
• Deployed in outbound sales and customer service scenarios

技能

• 自然语言处理
• 大语言模型训练与调优
• 多模态模型开发
• 团队管理与协作

Skills

• Natural Language Processing
• Large Language Model Training & Fine-tuning
• Multimodal Model Development
• Team Management & Collaboration

出版物

• Yongle Li, Zheng Zhang, 等. AtomTool: Empowering Large Language Models with Tool Utilization Skills. PRCV 2024.
• Yongle Li, Bo Liu, Sheng Huang, Zheng Zhang, Xiaotong Yuan, Richang Hong. Communication-Efficient and Personalized Federated LLM Fine-Tuning via Tri-Matrix Adaptation.
• 张俊祺、张正等：《Llama 大模型实践指南》电子工业出版社，2024
• Zijun Chen, Zheng Zhang, 等. Unveiling Uncertainty: Calibration of Multimodal LLMs.

Publications

• Yongle Li, Zheng Zhang, et al. AtomTool: Empowering Large Language Models with Tool Utilization Skills. PRCV 2024.
• Yongle Li, Bo Liu, Sheng Huang, Zheng Zhang, Xiaotong Yuan, Richang Hong. Communication-Efficient and Personalized Federated LLM Fine-Tuning via Tri-Matrix Adaptation.
• Junqi Zhang, Zheng Zhang, et al. “Llama Large Model Practice Guide” Publishing House of Electronics Industry, 2024
• Zijun Chen, Zheng Zhang, et al. Unveiling Uncertainty: Calibration of Multimodal LLMs.