Subscription fatigue isn't a UX problem — it's a behavioral design problem. Users don't need "easier cancellation" — they need visibility into subscription value before they hit the cancel button.
做了什么What I built
全程主导产品：12次用户访谈 → 从"取消流程"转向"健康仪表板"
在 Google AI Studio 设计交互流程，优化数据追踪指标
带领4人团队推进 MVP 开发（React Native）
Owned end-to-end product: 12 user interviews → pivoted from "cancel flow" to "health dashboard"
Designed interaction flows in Google AI Studio, optimized for metrics tracking
Leading 4-person team through MVP development (React Native)
交付物Deliverables
🎨
UI/UX Prototype — Google AI Studio
高保真原型 · 订阅健康仪表板交互流程
High-fidelity prototype · Subscription Health Dashboard interaction flow
Self-reflection isn't journaling — it's a behavioral design problem. Users don't need more text boxes. They need high-quality AI-assisted insights at the right moment to see their own thinking patterns.
做了什么What I built
设计7天周期系统 + 4天解锁机制，降低 time-to-value
通过 Vercel Functions 代理 API 调用——前端零密钥暴露
为 Reflect（模式+障碍）和 Rewire（心态+行动）分别设计提示词
Designed 7-day cycle system + 4-day unlock threshold to reduce time-to-value
Proxied API calls through Vercel Functions — zero API key exposure in frontend
Separate prompt design for Reflect (patterns + blocks) and Rewire (mindset + action)
Trustworthiness isn't a feature you add at the end — it's a framework you design into the product from day one. Most AI teams optimize for accuracy when users actually care about consistency.
做了什么What I built
将学术信任维度转化为产品团队可实际使用的评估指标
与初创公司合作，为 LLM 部署构建评估框架
产出完整操控分类体系，含风险评估和产品影响分析
Translated academic trust dimensions → operational product metrics teams can actually use
Partnered with startup to build evaluation framework for LLM deployment
Produced full manipulation taxonomy with risk assessment and product implications
AI in healthcare fails quietly. Benchmarks test best-case scenarios, but real patients give messy, ambiguous inputs — that's where models break. Edge cases reveal what averages hide.
做了什么What I built
文献综述 → 提取关键患者行为维度用于基准设计
创建并标注 3,000+ 基准数据点进行鲁棒性测试
论文已被 ICLR 2026 可信AI Workshop 录用（共同作者）
Literature review → extracted key patient behavior dimensions for benchmark design
Created + annotated 3,000+ benchmark datapoints for robustness testing
Co-authored paper accepted at ICLR 2026 Trustworthy AI Workshop
交付物Deliverables
📄
研究论文 · ICLR 2026 可信AI WorkshopResearch Paper · ICLR 2026 Trustworthy AI Workshop
At billion-impression scale, execution IS strategy. You can't "growth hack" your way through 500 stakeholders — creative problem-solving and systematic communication save chaos.
做了什么What I built
主导全程：受众分层 → KOL投放 → 效果追踪
跨中国/东南亚市场协调创意、媒体、数据团队
管理品牌方、代理商及500+ KOL的利益相关方沟通
Led end-to-end campaign: audience segmentation → KOL seeding → performance tracking
Coordinated creative, media, analytics teams across China/SEA markets
Managed stakeholder communication across brand, agency, and 500+ KOL partners