AI Product Manager
Agent & Data Strategy
Building AI-Native products at Baidu — Agent platforms, multi-scenario evaluation systems, and training-data strategy across 15+ business lines.
Currently leading Agent platforms, multi-scenario evaluation systems, and training-data strategy at Baidu — covering medical agents, virtual-companion agents, and 15+ business lines.
I practice an AI Native methodology: Agent as a first-class citizen, Specification-Driven, Workflow + Agent hybrid, Evaluation-First, Vibe Coding.
A compact case index: each row exposes the outcome first, then expands into the system logic behind it.
Agent-Native platform with claim-based task dispatching, pre-qualification gating, and heterogeneous role collaboration across annotator and QA agents.
抢注式任务分发 + 资格门控 + 异构角色协作，Plan-Act + Reflection 推理范式，Workflow + Agent 混合架构。
Five-stage evaluation pipeline: wiki, generation, refine, eval, analyse. Wiki-RAG enhanced Rubric generation plus Propose-Evaluate-Revise self-iteration.
医疗问诊 Agent 五阶段解耦评测 Pipeline，覆盖 50+ 临床病种，Cohen’s κ = 0.78。
Specification-Driven evaluation for virtual companion dialogue agents across WenXiaoYan, Shoubai, and in-car products.
-1/0/1 Likert scoring + 11-dimension Analytic Rubric; 21+ batch runs; label consistency 97%。
End-to-end automated pipeline for SFT / DPO / preference data across 15+ business lines, with Bad Case feedback loops as a standard correction mechanism.
线上 Bad Case 回流 → 归因 → 数据补充，面向概率性 Agent 产品建立标准纠错机制。
Full-stack HR operations platform for data-annotation teams: nine-dimension roster management, cross-project dispatch with approval flows and timelines, unified person-day performance metrics, and natural-language AI queries. React 18 + Express 5 + Prisma 6 + PostgreSQL 16, shipped via Docker Compose.
面向数据标注团队的项目人力看板与协同平台——花名册、跨项目调度审批、绩效与成长值统计、AI 自然语言查询、三级 RBAC 权限体系。React + Express + Prisma + PostgreSQL，Docker Compose 部署。
Three-layer evaluation for a micro-expression labeling system (spec v2.7): S0–S5 clip screening with hard filters and VLM quality gates, IAA-based spec stability, VLM-vs-human pipeline accuracy, and a label→generate→relabel semantic-fidelity loop, backed by a seven-factor ablation framework.
微表情标签体系三层评测——S0–S5 视频片段筛选（硬筛 + VLM 质量门）、规范稳定性（IAA）、pipeline 准确度（VLM vs 人工）与端到端语义保真闭环，配套 7 因子消融实验框架。
Professional tarot-reading MVP — a GSAP-driven shuffle-and-draw ritual, a rules engine computing elemental dignities and numeric arcs for free readings, two-tier paid AI readings via DeepSeek, and a complete international payment loop with Creem, Cloudflare Functions and KV.
专业塔罗阅读 MVP——GSAP 洗牌选牌仪式，基于元素尊卑与数字弧的规则引擎驱动免费解读，DeepSeek 双档位付费 AI 解读，Creem + Cloudflare Functions + KV 跑通完整海外收款链路。
A multi-agent system modeling real companies — AI agents communicate peer-to-peer through Feishu/Lark with LLM-planned discussion phases.
A knowledge base for LLM evaluation: methodologies, benchmarks, model comparisons, industry trends, and hands-on experience.
面向大模型评测工作的知识库：评测方法论、基准测试、模型对比、行业动态和评测实践经验。
Daily AI intelligence digest — automatically fetches trending AI projects from GitHub, performs deep analysis via LLMs, and delivers daily digest emails.
每日 AI 情报摘要系统：自动获取 AI 领域热点项目，通过大模型深度分析后发送每日情报邮件。
The original card grid is reframed as an operating model: specs define good work, agents execute open-ended work, workflows keep deterministic control, and evaluation closes the loop.