Wei Yang

Production AI, search, CMS, and personalization for industrial digital platforms.

• Built production RAG conversational AI and hybrid-search experiences with Azure OpenAI, Redis, and Weaviate
• Designed recommendation and offline ML pipelines for personalization, funnel optimization, and targeting
• Integrated partner/location CRM updates via MuleSoft into Solr and Elasticsearch with caching
• Developed licensed AEM CMS software for content authoring and publishing, integrating with search, PIM, marketing, and ecommerce platforms
• Translated business and UX requirements into modular, decoupled components; maintained data stores, documentation, and QA for reliable operations
• Hardened delivery with secure AI workflows, multilingual retrieval, and maintainable cross-system data flows

Deep learning and NLP engineering for financial text analysis.

• Built and maintained data-processing and training workflows for NLP experiments on financial text
• Modularized pipelines and PyTorch/ML helpers into reusable utils for shared research use
• Set up and operated remote Ubuntu GPU environments for deep learning training and evaluation
• Supported language-feature and text-classification studies with domain-adapted BERT models

AI content transformation — PDFs become semantic HTML and structured AEM components for enterprise publishing.

• Delivered PDF-to-semantic-HTML workflows that produce authored Adobe Experience Manager components
• Implemented an agentic LLM-vision pipeline to parse layout and generate reusable AEM components
• Added I18N translation support governed by brand voice and editorial guidelines
• Designed the tools to run on existing customer infrastructure without external processing servers

Agentic AI conversation platform — knowledge-grounded engagement with playbooks and tools for marketing-funnel optimization.

• Contributed to proof-of-concept development and architecture discussions for a platform that answers visitor questions from unified knowledge sources with citations
• Provided architecture recommendations for knowledge ingestion, retrieval, agentic playbooks, internal tools, conversation flows, and enterprise deployment
• Researched and optimized retrieval quality, response grounding, agentic routing, playbook and tool behavior, and marketing-funnel performance

Real-time AI search overview — hybrid retrieval across heterogeneous enterprise sources.

• Built real-time search overview and autosuggest with live-query caching for low-latency responses
• Engineered multilingual indexing across 10+ sources with hybrid search, reranking, routing, and iterative retrieval
• Added reporting pipelines and feedback loops to improve overview relevance and quality over time
• Separated overview generation from chat so each surface could optimize latency, ranking, and UX independently

Account-based marketing — website experiences tailored to target accounts and buying contexts.

• Extended recommendation workflows from product-level targeting to account-aware journeys
• Aligned page content and offers with account segments and prior interaction signals
• Connected personalization outputs to account segments used across marketing surfaces

Agentic conversational AI — enterprise knowledge retrieval with measurable production usage.

• Built an agentic chatbot with Azure OpenAI, Redis, and Weaviate, integrating internal company tools and APIs through function calling
• Designed multi-turn Redis memory and agentic function-calling workflows for tools and APIs
• Implemented security guardrails including input sanitization, content filtering, and PII masking
• March 2026 usage: approximately 110K unique sessions across the tracked search, retrieval-assistant, and fault-code applications
• Delivered the website chat entry point for AI-powered support and quick access to trusted information

Recommendations and offline ML — behavioral data turned into personalized product experiences.

• Built user-to-item and item-to-item recommendation pipelines with rolling cache and Akamai CDN
• Designed offline ML pipelines for interaction modeling, personalization, and user classification
• Improved sales-funnel conversion and marketing targeting through customized product experiences
• Delivered discovery surfaces such as “Based on your views,” preferred products, and related content

AI-assisted technical support — fault-code lookup across industrial product documentation.

• Built AI-assisted lookup of condition, event, and fault codes across industrial product families
• Connected product-family selection and code queries to Technical Documentation Center retrieval
• Delivered production troubleshooting access for drives, controllers, and motion systems

Content-quality scoring — ML signals that accelerate marketing-funnel decisions.

• Built a backend analytics platform that scores content quality for funnel optimization
• Developed ML workflows to evaluate content effectiveness and prioritize conversion-focused changes
• Connected scoring outputs with marketing and content workflows for faster data-driven decisions

CRM-to-search data integration — partner and location records made reliably discoverable.

• Designed 7+ MuleSoft API workflows for incremental CRM delta updates of partner accounts and locations
• Integrated multi-source records into Lucidworks Solr and Elasticsearch search with cache optimization
• Improved reliability of partner and location data for searchable enterprise applications
• Continues in active maintenance and enhancement for production partner-locator experiences

Assessment questionnaires in AEM — first-party preference capture for cybersecurity readiness.

• Tailored AEM questionnaire components with category support, reporting, and dashboard presentation
• Supported a Cybersecurity Preparedness Assessment that produces a NIST-based score and customized report
• Improved structured authoring and multi-channel publishing for assessment-driven content workflows

On-site feedback loop — Hotjar signals wired into AEM-managed pages.

• Integrated Hotjar with Adobe Experience Manager to capture on-site feedback and behavioral signals
• Connected feedback collection to AEM pages so teams could identify UX friction and prioritize fixes
• Established an early feedback loop that informed later personalization and content-scoring work

Awarded the Provost’s Scholarship.

National Second Prize in the National Mathematical Modeling Contest; Provincial First Prize.

NLP and deep learning for tweet sentiment and market indicators.

• Cleaned and analyzed 20M+ related tweets with sentiment features at second-level granularity
• Engineered MACD/RSI-style indicators and back-tested multiple deep learning models for portfolio return

Job-description NLP with clustering and topic modeling.

• Scraped and structured Indeed job-description text; filtered noise with K-means skillset clusters
• Analyzed skill trends with topic modeling and aspect-extraction neural networks

Collaborative filtering on user, item, and interaction data.

• Built a recommender engine with memory-based and model-based collaborative filtering on JD.com interaction data
• Evaluated on 7,000 user-item samples and reached 77% Top-10 accuracy

Full-stack web app for furniture and rental information exchange.

• Led a four-person team to build a Node.js/Express app with MongoDB schemas and RESTful CRUD APIs
• Implemented authentication, search, comments, and dashboard features for listing exchange

City vibrancy modeling from multi-source public data.

• Built a vibrancy index from Google POIs, Instagram, Zillow, and BLS data with Pandas feature engineering
• Fine-tuned XGBoost and Keras models to predict U.S. city prosperity trends

面向工业数字平台的生产级 AI、搜索、CMS 与个性化。

• 使用 Azure OpenAI、Redis 与 Weaviate 构建生产级 RAG 对话式 AI 与混合搜索体验
• 设计用于个性化、漏斗优化与定向的推荐与离线 ML 流水线
• 通过 MuleSoft 将合作伙伴/地点 CRM 更新集成到 Solr 与 Elasticsearch，并配合缓存
• 开发可授权的 AEM CMS 软件，用于内容创作与发布，并与搜索、PIM、营销与电商平台集成
• 将业务与 UX 需求转化为模块化、解耦组件；维护数据存储、文档与 QA，保障可靠运维
• 通过安全 AI 工作流、多语言检索与可维护的跨系统数据流强化交付

面向金融文本分析的深度学习与 NLP 工程。

• 构建并维护面向金融文本 NLP 实验的数据处理与训练工作流
• 将流水线与 PyTorch/ML 辅助工具模块化为可复用 utils，供研究共享使用
• 搭建并运维远程 Ubuntu GPU 环境，用于深度学习训练与评估
• 以领域自适应 BERT 模型支持语言特征与文本分类研究

AI 内容转换——将 PDF 转为语义 HTML 与结构化 AEM 组件，用于企业发布。

• 交付 PDF 到语义 HTML 工作流，生成可编辑的 Adobe Experience Manager 组件
• 实现智能体 LLM-vision 流水线以解析版面并生成可复用 AEM 组件
• 增加受品牌语气与编辑规范约束的 I18N 翻译支持
• 设计工具在客户现有基础设施上运行，无需外部处理服务器

智能体 AI 对话平台——知识 grounding 互动，配合 playbook 与工具优化营销漏斗。

• 参与概念验证开发与架构讨论，支持基于统一知识源并附引用的访客问答
• 就知识摄入、检索、智能体 playbook、内部工具、对话流程与企业部署提供架构建议
• 研究并优化检索质量、回答 grounding、智能体路由、playbook 与工具行为，以及营销漏斗表现

实时 AI 搜索概览——跨异构企业数据源的混合检索。

• 构建实时搜索概览与自动建议，含实时查询缓存以实现低延迟响应
• 工程化跨 10+ 数据源的多语言索引，含混合搜索、重排序、路由与迭代检索
• 增加报表流水线与反馈闭环，持续改进概览相关性与质量
• 将概览生成与聊天分离，使各界面可独立优化延迟、排序与 UX

基于账户的营销（ABM）——按目标账户与购买情境定制网站体验。

• 将推荐工作流从产品级定向扩展到账户感知旅程
• 使页面内容与优惠与账户细分及既往交互信号对齐
• 将个性化输出连接到跨营销界面使用的账户细分

智能体对话式 AI——具备可衡量生产使用量的企业知识检索。

• 使用 Azure OpenAI、Redis 与 Weaviate 构建智能体聊天机器人，通过函数调用集成内部公司工具与 API
• 设计多轮 Redis 记忆与面向工具/API 的智能体函数调用工作流
• 实现安全护栏，包括输入净化、内容过滤与 PII 掩码
• 2026 年 3 月使用量：在所追踪的搜索、检索助手与故障代码应用中约有 110K 独立会话
• 交付网站聊天入口，提供 AI 驱动支持与可信信息的快捷访问

推荐与离线 ML——将行为数据转化为个性化产品体验。

• 构建用户到物品与物品到物品推荐流水线，含滚动缓存与 Akamai CDN
• 设计用于交互建模、个性化与用户分类的离线 ML 流水线
• 通过定制化产品体验提升销售漏斗转化与营销定向
• 交付如“Based on your views”、偏好产品与相关内容等发现界面

AI 辅助技术支持——跨工业产品文档的故障代码查询。

• 构建跨工业产品系列的条件、事件与故障代码 AI 辅助查询
• 将产品系列选择与代码查询连接到技术文档中心检索
• 为驱动器、控制器与运动系统交付生产级排障访问

内容质量评分——加速营销漏斗决策的 ML 信号。

• 构建用于漏斗优化的内容质量评分后端分析平台
• 开发 ML 工作流以评估内容效果并优先推进转化导向改动
• 将评分输出与营销及内容工作流连接，加快数据驱动决策

CRM 到搜索数据集成——使合作伙伴与地点记录可靠可发现。

• 设计 7+ 个 MuleSoft API 工作流，用于合作伙伴账户与地点的 CRM 增量更新
• 将多源记录集成到 Lucidworks Solr 与 Elasticsearch 搜索，并优化缓存
• 提升合作伙伴与地点数据在可搜索企业应用中的可靠性
• 持续对生产级合作伙伴定位体验进行主动维护与增强

AEM 中的评估问卷——面向网络安全准备的第一方偏好采集。

• 定制带分类支持、报表与仪表盘展示的 AEM 问卷组件
• 支持生成基于 NIST 评分与定制报告的网络安全准备评估
• 改进评估驱动内容工作流的结构化创作与多渠道发布

站内反馈闭环——将 Hotjar 信号接入 AEM 管理页面。

• 将 Hotjar 与 Adobe Experience Manager 集成，采集站内反馈与行为信号
• 将反馈采集连接到 AEM 页面，便于团队识别 UX 摩擦并优先修复
• 建立早期反馈闭环，为后续个性化与内容评分工作提供依据

获颁教务长奖学金（Provost’s Scholarship）。

全国大学生数学建模竞赛全国二等奖；省级一等奖。

面向推文情感与市场指标的 NLP 与深度学习。

• 清洗并分析 2000 万+ 相关推文，以秒级粒度提取情感特征
• 工程化 MACD/RSI 类指标，并对多种深度学习模型进行回测以评估组合收益

结合聚类与主题建模的职位描述 NLP。

• 抓取并结构化 Indeed 职位描述文本；用 K-means 技能集聚类过滤噪声
• 用主题建模与方面抽取神经网络分析技能趋势

基于用户、物品与交互数据的协同过滤。

• 基于京东交互数据，用基于记忆与基于模型的协同过滤构建推荐引擎
• 在 7,000 个用户-物品样本上评估，达到 77% Top-10 准确率

面向家具与租房信息交换的全栈 Web 应用。

• 带领四人团队构建 Node.js/Express 应用，含 MongoDB schema 与 RESTful CRUD API
• 实现认证、搜索、评论与仪表盘功能，支持信息发布与交换

基于多源公共数据的城市活力建模。

• 用 Google POIs、Instagram、Zillow 与 BLS 数据构建活力指数，并用 Pandas 做特征工程
• 微调 XGBoost 与 Keras 模型，预测美国城市繁荣趋势