Jiawei Lu

一个以做东西为主的实习，同时碰了几个 AI 医疗产品：远程患者监测、主动外呼的 AI 语音智能体、AI 营销内容流水线，每个都负责了一大块。

A build-first internship across several AI-driven healthcare products — remote patient monitoring, a proactive AI voice agent, and an AI marketing content pipeline — owning substantial pieces of each end-to-end.

• 和团队一起搭远程患者监测和照护协调平台，前端 React/TypeScript，后端 Python，云基础设施，重点补自动化测试。
• 参与用药依从性外呼的 AI 语音智能体，做 LLM 编排，以及一个对比不同模型后端回答质量的基准工具。
• 帮着上线 AI 营销内容流水线：生成、合规审核、多渠道排期，还给一个医疗相关客户做了带完整审计记录的扩展。
• 经常同时开多个 git worktree，功能、基础设施修复、稳定性加固并行推进。

• Co-built a remote patient-monitoring and care-coordination platform across the full stack (React/TypeScript frontend, Python backend, cloud infrastructure) with a strong emphasis on automated test coverage.
• Contributed to a proactive AI voice agent for medication-adherence outreach: LLM orchestration and a benchmark harness for comparing model backends on response quality.
• Helped ship an AI-powered marketing content pipeline — generation, compliance review, multi-channel scheduling — including an extension for a healthcare-adjacent client with a full audit-trail requirement.
• Regularly ran multiple parallel git worktrees to move across concurrent feature work, infra fixes and reliability hardening.

在自研大模型的数据侧，做中文对话数据，语料来自相声录音。

On the data pipeline side of in-house LLM development, focusing on Chinese conversational data sourced from crosstalk (相声) recordings.

• 设计三级过滤流水线（正则规则 → LLM 判别 → 人工抽检），一万多条原始对话筛到约六千条训练样本。
• 写 few-shot 样例，让 LLM 改写信息密度低的对话片段。
• 搭 GPT / Claude / Qwen 多模型路由，异步调用快约 5 倍，本地并行做基准测试快约 10 倍。
• 跨模型做 prompt 位置消融，发现 System 和 User 角色注入的最优位置因模型而异，直接影响了流水线设计。
• 清洗后的数据让内部推理 / 语言基准涨 6–7 分，过了上线门槛。

• Designed a three-stage filtering pipeline (rule-based regex → LLM-as-a-Judge → manual spot-check) to turn 10K+ raw dialogues into ~6K gold-standard training samples.
• Developed few-shot examples to guide LLM-assisted rewriting of low-density dialogue segments.
• Built a multi-model routing pipeline (GPT / Claude / Qwen) with async API execution (~5× speedup) and local parallelisation for benchmarking (~10×).
• Ran prompt-placement ablations across models; optimal System-role vs. User-role injection varied by model, which shaped the pipeline design.
• The curated dataset contributed to a 6–7 point gain on internal reasoning/language benchmarks, clearing the bar for consumer deployment.

销售团队要跟品牌客户（主要是健康类）解释广告有没有用、和竞品比在什么位置。我搭了支撑这些对话的数据和分析。

The sales team needed to show brand clients — mostly in health & wellness — whether their ad spend was working and how they stacked up against competitors. I built the data and analysis behind those conversations.

• 用 Selenium 自动化内部分析平台的取数，从半天缩到 30 分钟以内。
• 用 Excel 给十多个隐形眼镜品牌做多维分析，按产品线、时间段、投放位置看回搜率和转化率。
• 做匿名化的竞品对标报告，客户能看同行表现但看不到是谁。
• 给销售和品类经理出建议，四个重点客户基本全部采纳，影响了后续预算。
• 参与投后复盘，准备复盘 deck 的数据部分。

• Automated data retrieval from the internal analytics platform via Selenium, cutting pull time from half a day to under 30 minutes.
• Built Excel-based multi-dimensional analyses tracking return-search and conversion rates across product lines, periods and placements for 10+ contact lens brands.
• Produced anonymised competitive benchmarking reports.
• Delivered recommendations to sales and category managers; 4 key accounts adopted nearly all suggestions.
• Prepared the data components of post-campaign recap decks.

• 分析经销商层面的销售数据，找出表现差的门店，标记需要正式干预的。
• 警告函分发用邮件合并自动化，每个高峰周期十到十五封，从两天缩到一个下午。

• Analysed dealer-level sales data to identify underperforming stores and flag those requiring formal intervention.
• Automated warning-letter distribution via mail merge, reducing turnaround from ~2 days to a single afternoon for ~10–15 letters per peak cycle.

让小模型写出跑得快的 SQL，不只是写对。在 BIRD 基准上两阶段微调 Qwen3-4B。

Can a small model generate not just correct SQL, but efficient SQL? Two-stage alignment of Qwen3-4B on the BIRD benchmark.

• 先用高效 / 低效 SQL 配对做 DPO，再用带延迟奖励的执行感知强化学习（GRPO 风格）。
• BIRD Mini-Dev 上执行准确率和 R-VES 比基线高 7%。
• 对查询结构和优化模式做了定性错误分析。

• DPO on curated efficient-vs-inefficient SQL pairs, then execution-aware RL (GRPO-style) with latency-based rewards.
• 7% improvement in execution accuracy and R-VES over baseline on BIRD Mini-Dev.
• Qualitative error analysis on query structure and optimisation patterns.

• 三年里从宣传干事到宣传部长再到副主席。
• 一个月两场活动，人头、物料、时间表、对外沟通从头管到尾。
• 带 25 人的宣传组，负责公众号和其他平台的内容。
• 上海马拉松志愿项目从招募到复盘的整个宣传周期。

• Publicity Officer → Head of Publicity → Vice President over three consecutive years.
• Co-led ~2 events per month: headcount, materials, schedules and external stakeholder communication.
• Managed a 25-person publicity team producing content for the WeChat Official Account and other platforms.
• Led the full publicity cycle for the Shanghai Marathon volunteer programme, from recruitment to post-event recap.