小鹏汽车GPU tools高级/资深/专家工程师

1. Architect and implement agentic workflows that plan, reason, call tools/APIs, and coordinate with humans or other agents. 2. Select, extend, or build frameworks (e.g., LangChain, AutoGen, CrewAI, MetaGPT, LangGraph) to accelerate delivery while avoiding vendor lock-in. 3. Own the MLOps lifecycle: data collection, evaluation harnesses, safety filters, CI/CD, and observability for deployed agents. 4. Integrate enterprise systems & data sources (REST/GraphQL, Kafka, vector databases, Kubernetes) so agents can act on real business objects. 5. Mentor and review code for junior engineers; drive best practices in prompt engineering, evaluation, and secure coding. 6. Research emerging techniques (toolformer, self-reflection, role specialization) and translate findings into the product roadmap.

【为什么加入我们】 ● 我们是谁：阿里国际数字商业集团下Alibaba.com公司（简称阿里国际站），目前为全球最大的在线跨境B2B贸易平台。 ● 真实的复杂场景： 区别于简单的对话，B2B 贸易涉及高客单价、长决策链路（20+ 节点）及深度供应商研究，是 Agentic Workflow 落地最具价值的试验田。 ● 顶尖科研环境： 团队成员来自国内外名校，拥有硅谷 Office 协同，支持前沿算法探索。 ● 充足算力支持： 自有国内外机房，GPU 资源充足，拒绝“无米之炊”。 【岗位职责】 作为核心算法专家及团队负责人，你将主导 AI 搜索与 Agent 体系的架构演进： 1. Agent 架构设计： 负责 B2B 买家 Agent 的核心引擎开发，包括但不限于 Planning（任务规划）、Long-term Memory（长期记忆）、Tool-use（工具调用）及 Reflection（自我反思） 等模块。 2. 下一代搜索系统： 构建多轮交互下的多模态搜索意图理解。针对超长图文、PDF 采购文件等复杂私有知识，设计高性能的 RAG（检索增强生成）与多模态匹配系统。 3. 算法团队管理： 带领算法团队进行技术攻关，负责从 0 到 1 的算法落地与迭代，建立高效的团队协作与技术梯队。 4. 业务洞察与驱动： 利用 AI 分析海量用户行为，挖掘跨境贸易中的低效环节，通过算法创新定义新的产品形态（AI-Native）。

• Lead the software development in C/C++, Python, and in GPU languages such as CUDA, ROCm, or Triton• Analyze metrics and identify opportunities based on offline and online testing, develop and deliver robust and scalable solutions.• Work with cutting-edge hardware stacks and a fast-moving software stack to deliver best-of-class inference and optimal cost.• Engage with key partners to understand and implement inference and training optimization for state-of-the-art LLMs and other models.

• Architect Performance Tooling: Develop infrastructure tools/libraries for GPU performance analysis, visualization, and automated workflows used across GPU SW/HW development life cycle.   • Unlock Architectural Insights: Analyze GPU workloads to identify bottlenecks and define new hardware profiling features that enhance perf debug and profiling capabilities.  • AI-Powered Automation: Build AI/ML-driven tools to automate performance analysis, generate perf optimization guidance, and improve user experience of profiling infrastructure.  • Cross-Stack Collaboration: Partner with kernel developers, system software teams, and hardware architects to support performance study, improve CUDA software stack, and co-design performance-centric solutions for current and next-generation GPU architecture