几件事 不同节奏 同时进行
A few things different speeds at the same time
刚做完的：在 xHealth 做健康 AI 对话，让它记住患者之前说过什么，有高风险信号就推给医生
Just finished: a health AI chat agent at xHealth. It remembers what a patient said before and flags high-risk signals to a doctor
一直在学的：统计模型，大模型的训练数据从哪来，数据流水线一般在哪步出问题
Always learning: statistical models, where LLM training data comes from, and which step of a data pipeline usually breaks
一直没停的：插花。闲下来的周末不多，有的话就去剪枝
Never stopped: flowers. Not many free weekends, but the ones I get go to cutting stems
做过的事 六段
Things I've done six of them
xHealth Group
软件工程实习生SUMMER ENGINEERING INTERN十周，从想法做到 Demo Day，一个健康 AI 对话产品从头到尾我负责。它要记住患者长期的情况，做症状追踪和用药提醒，高风险信号推给医生复核，输出要能给临床用、能追溯。中间还和两个同事搭了个内部 AI 平台
Ten weeks, idea to Demo Day, one health AI chat product that was mine end to end. The agent keeps a patient's long-term context for symptom tracking and medication reminders, and pushes high-risk signals to a clinician for review. Outputs have to be usable in clinic and auditable. Also built an internal AI platform with two colleagues
阶跃星辰Stepfun
大模型数据工程实习生LLM DATA ENGINEERING INTERN做中文对话数据，语料是相声录音。三级过滤（正则 → LLM 判别 → 人工抽检）把一万多条原始对话筛到六千条左右。多模型路由 GPT / Claude / Qwen 加异步调用快了 5 倍，本地并行 10 倍。清完的数据让内部基准涨了 6–7 分，过了上线的线
Chinese conversational data, sourced from crosstalk recordings. A three-stage filter (regex → LLM-as-a-Judge → manual spot-check) took 10K+ raw dialogues down to about 6K. Multi-model routing across GPT / Claude / Qwen with async calls ran about 5× faster, local parallelisation about 10×. The cleaned set moved internal reasoning and language benchmarks 6–7 points, enough to ship
小红书Red Note
快消部门数据分析实习生DATA ANALYST INTERN, FMCG销售要跟品牌客户解释广告有没有用、和竞品比在什么位置。我用 Selenium 把取数自动化了，半天变 30 分钟；给十几个隐形眼镜品牌做了回搜率和转化率的分析。竞品对标做了匿名化，客户能看同行但看不到是谁。四个重点客户基本全采纳了
Sales had to explain to brand clients whether their ads worked and where they sat against competitors. I automated the data pulls with Selenium, half a day down to under 30 minutes, and built return-search and conversion analyses for 10+ contact lens brands. Competitor benchmarks were anonymised so clients could see peers without names. Four key accounts took nearly all the recommendations
大陆集团Continental Tires
运营分析实习生OPERATIONS ANALYST INTERN看经销商的销售数据，找出表现差的门店，标记要正式干预的。警告函用邮件合并自动发，一个高峰周期十来封，两天缩到一个下午
Dealer-level sales analysis to find underperforming stores and flag the ones needing formal intervention. Mail-merge automation cut warning-letter turnaround from about two days to an afternoon, 10–15 letters per peak cycle
NL to Performant SQL
COLUMBIA · NLP让小模型写出跑得快的 SQL，不只是写对。在 BIRD 上两阶段微调 Qwen3-4B-Instruct：先用高效 / 低效 SQL 配对做 DPO，再用带延迟奖励的 GRPO。Mini-Dev 上执行准确率和 R-VES 比基线高 7%
Getting a small model to write SQL that runs fast, not just SQL that is correct. Qwen3-4B-Instruct fine-tuned on BIRD in two stages: DPO on efficient-vs-inefficient SQL pairs, then GRPO with latency-based rewards. 7% over baseline on execution accuracy and R-VES on Mini-Dev
上海交大 电院志愿者协会SJTU SEIEE Volunteer Association
副主席 · 宣传部长VICE PRESIDENT · HEAD OF PUBLICITY两年，从在活动上拍照到和别人一起管一个 130 人的社团。宣传干事 → 宣传部长 → 副主席，一个月两场活动，带 25 人的宣传组。上海马拉松志愿项目从招募到复盘的宣传都是我们做的
Two years, from taking photos at events to co-running a 130-person organisation. Publicity Officer → Head of Publicity → Vice President, about two events a month, a 25-person publicity team, and the full publicity cycle for the Shanghai Marathon volunteer programme, recruitment to recap
花期 慢一点的部分
In Bloom the slow part
花没办法加速，剪一枝之前要看整体，剪完再看一遍。做这个的时候脑子里只有这一件事
Flowers don't speed up. You look at the whole thing before each cut, then look again after. While I'm doing this it's the only thing in my head
FLORAL 01作品 01
Arrangement 01
FLORAL 02作品 02
Arrangement 02
FLORAL 03作品 03
Arrangement 03
有想聊的 写邮件就行
Want to talk email works
在找 2027 年之后数据 / AI 工程方向的机会，合作也欢迎。邮件会回，快慢看那天状态
Looking for data / AI engineering roles from 2027. Collaborations welcome too. I do reply to email, how fast depends on the day