把想法Make 做成ideas 作品。real.

负责 AI 与创新中心，从 0 到 1 搭建面向行业从业者的 AI 工作助手，以及它背后的数据基座、智能体和应用产品矩阵。

• 写下团队的 AI 战略架构：四层结构、「三焊点」护城河、与基座模型保持正夹角、有界的递归式自我改进
• 推动 AI 从「读」走向「写」：让智能体被授权改变业务状态，并定义写操作的授权与审计模式
• 用一个北极星指标统一十几条项目线的汇报口径，每个产品只认领漏斗的一跳

I head the AI & Innovation Center and built, from zero, an AI work assistant for industry practitioners along with the data foundations, agents and application matrix behind it.

• Authored the team's AI strategy architecture: four layers, the "three welds" moat, a positive angle to foundation models, bounded recursive self-improvement
• Moved the AI from reads to writes: agents authorized to change business state, with a defined authorization and audit pattern
• Unified a dozen workstreams under one north-star metric, each product claiming exactly one funnel step

工作六年后回到学校，系统补上深度学习这一层：深度神经网络、自然语言处理、计算机视觉、强化学习、AI Agent。课程项目做了两个能跑起来的东西：一个通过聊天机器人实时交互的民宿推荐系统（NeuralCF），一个用 CNN + BERT 做的多模态鸟类识别与问答系统（部署在 Telegram）。

Back to school after six years of work to properly learn the deep-learning layer: deep neural networks, NLP, computer vision, reinforcement learning, AI agents. Two course projects that actually ran: a real-time BnB recommender driven by a chatbot (NeuralCF), and a multimodal bird identification and Q&A system built on CNN + BERT, deployed as a Telegram bot.

负责中国西部航空业务的机票订单数据分析，为一块年销售约 200 亿美元的业务提供判断和决策建议，帮团队把订单量和销售额保持在每年 15% 的增速。2019 年起同时孵化并负责集团的公务机包机业务：一个面向全球用户的在线包机预订平台，从产品、增长、对外谈判到团队管理都由我来跑。这段经历是我后来做「垂直行业 AI」的底子：先懂业务和数据，再谈模型。

• 「低效航班」专项：对表现不佳的航班做精细化数据运营，把区域年增速拉高 1.2 个百分点，多卖 50 万张票，增加约 7000 万美元营收
• 每年为航司和机场输出 30+ 份数据分析与优化方案（开航、促销等），为合作方带来最高约 1000 万美元的年增量收益
• 公务机业务：用深度学习做客户需求预测来指导航司调度，为部分客户降低最高 50% 成本；用数据挖掘搭精准营销系统，找出贡献未来 20% 订单量的潜在用户

Covered air-ticket order data for the Western China aviation business, a unit with roughly $20B in annual sales, providing the judgment and recommendations that kept order volume and revenue growing about 15% a year. From 2019 I also incubated and ran the group's business-jet charter business, an online booking platform for a global audience, owning product, growth, external negotiations and the team. This is where my approach to vertical AI comes from: understand the business and the data first, then talk about models.

• "Underperforming flights" program: fine-grained data operations on weak routes lifted regional annual growth by 1.2 points, about 500,000 extra tickets and roughly $70M in revenue
• Delivered 30+ data analyses and optimization plans a year for airlines and airports (route launches, promotions), worth up to about $10M in annual incremental value to partners
• Business jets: a deep-learning demand forecaster to guide operators' scheduling, cutting costs by up to 50% for some customers; a data-mining marketing system that identified users contributing over 20% of future orders

参与集团的「机+酒」打包推荐项目：给买机票的用户推荐最合适的酒店。跑了 50 多个城市的酒店实地评估，用聚类和基于 LBS 的可视化找业务规则，和算法团队一起打磨推荐模型，把机酒组合的推荐成交率从 2% 提到 20%。毕业后因此进入高级总监直属团队。

Worked on the group's "Flight + Hotel" package recommender, matching air-ticket buyers with the right hotel. Evaluated hotels on site in 50+ cities, used clustering and LBS-based visualization to surface business rules, and refined the model with the algorithm team, taking recommendation-to-booking from 2% to 20%. Joined the senior director's team on graduation as a result.

负责咪咕音乐搜索引擎千万级日查询的数据分析与统计，基于用户搜索日志建立搜索期望评分模型；写 Python 脚本采集并挖掘竞品的异常数据，优化搜索、纠错与联想引擎，搜索期望值提升近 5%。

Analyzed tens of millions of daily queries for the Migu Music search engine, built a search-expectation scoring model from user logs, and wrote Python scripts to collect and mine anomalies from competing products to improve search, error tolerance and suggestions, lifting search expectation by nearly 5%.

用 Presto 做订单与用户画像的数据提取分析，为司机和乘客设计激励方案。司机激活项目里按行为给司机分层、匹配激励策略（比如演唱会散场时奖励兼职司机接单），激活转化率提升 35%，节省约 100 万美元激励成本；用 Tableau 和 matplotlib 把可视化报表的产出效率提高一半。

Pulled and analyzed order and rider-profile data in Presto to shape incentive schemes for drivers and passengers. In the driver-activation project, segmented drivers by behavior and matched incentives (such as rewarding part-time drivers to pick up after concerts), raising activation conversion 35% and saving about $1M in incentives; halved the effort of producing visual reports with Tableau and matplotlib.