长鑫存储设计分析工程师 | Design & Silicon Analysis Engineeer(J15916)

Team Introduction: Dedicated to building an industry-leading large-model dialogue system, the team serves hundreds of millions of daily active users, with application scenarios covering the entire Douyin e-commerce ecosystem. This includes core business scenarios such as platform customer service, platform merchant service, merchant customer service, influencer customer service, and innovative intelligent shopping guides. Through continuous technological innovation and optimization, the team has successfully established a complete intelligent dialogue solution, delivering significant efficiency improvements and user experience enhancements to e-commerce operations. Research Objectives: Develop an LLM-based customer service chatbot for TikTok and Douyin E-commerce, enabling intelligent customer service interactions. The LLM will handle the entire user inquiry process, including request clarification, solution negotiation, and execution. Necessity: LLM's strong conversational and reasoning abilities make it especially suitable for intelligent customer service, capable of potentially reaching the service standards of excellent human representatives. Research Content: Design a multi-agent framework based on LLM, integrating planning-agent, reply-agent, and tool-agent. Each agent will specialize in different functions, working collaboratively to manage the complete service process—from issue identification and solution negotiation to solution implementation and feedback. 1) Reply-agent ensures the proposed solutions comply with platform policies and service guidelines, avoids excessive improvisation or hallucinations, and maintains smooth communication and negotiation with the user. 2) Planning-agent identifies user demands and problem scenarios, sourcing relevant service guidelines and constraints as well as recognizing risk scenarios. 3) Tool-agent validates the legality of tool usage, accurately interprets the results from tool interactions, and manages execution dependencies of various actions. Research Challenges: Compliance with service guidelines: Ensuring the chatbot's solutions adhere to platform service guidelines (such as available refund within xx days of parcel arrival and coupon limits per user per week). Dynamic feedback adaptation: Static adherence to service rules and providing fixed solutions can limit the flexibility of reply-agents, preventing them from acting like excellent human customer service representatives. By enabling reply-agents to interact in real-time with their environment, considering user's behavioral trends, demands expressed during inquiries, and feedback on proposed solutions, personalized service can be provided. This approach fosters adaptive responses and progressive services and solutions, closely mirroring the flexibility and excellence of human customer service. Self-reflection: Employing LLM's capabilities to understand, analyze, and evaluate its own behavior, fostering self-supervision and decision refinement through reflection on outputs, particularly with complex and ambiguous tasks. Complex image processing: Handling scenarios involving numerous complex images (including shipping order photos, bank transaction screenshots, images of damaged goods received, and seller qualification certifications). These images contain key information crucial to enhancing the chatbot's problem resolution capabilities. 团队介绍： 智能对话团队，致力于打造业界领先的大模型对话系统。团队服务的日活用户超过数亿，应用场景覆盖抖音电商全链路，包括平台客服、平台商服、商家客服、达人客服，以及创新的智能导购等核心业务场景，通过持续的技术创新和优化，成功构建了一套完整的智能对话解决方案，为电商业务带来了显著的效率提升和用户体验改善。 课题目标： 构建基于LLM的电商客服机器人（Chatbot），服务TikTok和抖音电商智能客服场景，由LLM完成一次用户进线的完整接待过程，包括诉求澄清、方案协商、方案执行等阶段。 必要性： LLM具有强大的对话和推理能力，智能客服是LLM能够发挥价值的最典型场景，有机会能够达到匹配优秀人工客服的服务能力。 课题内容： 设计一个基于LLM 的 multi-agent framework，将 planning-agent、reply-agent、tool-agent 集成到一起，每个 agent 负责不同能力，互相协同，完成从问题定位、方案协商，到方案执行、结果反馈等服务全流程。reply agent 需要确保给用户提供的方案是符合平台的相关政策和service policy的，不自行过度发挥、不出现幻觉，顺滑的完成和用户的沟通协商过程；planning agent 完成定位用户诉求和问题场景，以便从外部获取该场景的服务准则和约束，如何识别风险场景；tool agent 需要确保工具调用的合法性、接收和解析工具调用的返回结果，另外一些动作的执行存在前后依赖的问题。 课题挑战： 1、遵循服务准则：如何确保方案Chatbot提供的方案是follow平台服务准则的，例如到货xx天之内可以申请退款、同一用户一星期内最多发送xx额度的优惠券； 2、感知环境反馈：reply agent如果只能死板的follow当前场景服务准则，提供一层不变的方案，是无法像优秀客服一样做到灵活变通的。让Agent能够实时的和环境打通，通过结合当前用户进线前的行为动线、进线后表达的诉求和用户对 agent 提供方案的反馈，为用户提供个性化的服务，对用户的实时反馈有响应，像优秀客服一样能随机应变，递进式的提供服务和解决方案； 3、进行自我反思：利用LLM理解、分析和评价其自身的行为，使LLM能够自我监督，通过对自身输出的反思，改进其所做的决策，以便在处理复杂、有歧义的任务时，能有更好的表现； 4、复杂图片理解：电商场景存在大量复杂的图片，包括运费订单实拍图、银行流水截图、买家收货缺件破损的、商家各类资质证明等，这类图片往往包含重要的信息，对提升Chatbot解决能力非常重要。

Team Introduction: The ByteDance Recommendation Architecture Team is responsible for the design and development of the recommendation system architecture for ByteDance's related products. It ensures the stability and high availability of the system, optimizes the performance of online services and offline data streams, resolves system bottlenecks, and reduces cost overheads. The team also abstracts the common components and services of the system, builds the recommendation middle - office and data middle - office to support the rapid incubation of new products and enable ToB services. 团队介绍： 字节跳动推荐架构团队，负责字节跳动旗下相关产品的推荐系统架构的设计和开发，保障系统稳定和高可用；负责在线服务、离线数据流性能优化，解决系统瓶颈，降低成本开销；抽象系统通用组件和服务，建设推荐中台、数据中台，支撑新产品快速孵化以及为ToB赋能。 课题背景： 在当今数字化时代，推荐系统已成为众多领域（如电商、信息资讯等）实现个性化服务、提升用户体验和竞争力的关键技术。然而，随着技术的不断发展和业务场景的日益复杂，推荐系统面临着诸多严峻挑战。 一方面，推荐系统自身的复杂性急剧增加。大量推荐策略不断演进迭代，且系统状态动态变化，但缺乏有效手段自动跟踪评估策略有效性并下线低 ROI 策略，导致系统存在较多低效策略。同时，推荐系统依赖多种基础组件，其复杂负载模型给底层组件参数配置和性能调优带来巨大困难，日常开发迭代中的问题排查等工作消耗大量人力，亟需提升开发效率、降低人力成本。 另一方面，随着电商行业等领域的激烈竞争，传统推荐系统在多样性、创新性和个性化方面的短板愈发凸显，难以满足用户日益增长的多元需求。生成式人工智能技术虽带来新突破，但在实际应用中面临成本效率、全域数据协同、数据隐私与安全以及技术变革应对等诸多难题。 此外，随着大模型的快速发展，推荐系统对用户行为序列数据的存储和质量要求不断提高，数据质量对模型性能的影响愈发关键。同时，模型规模的扩大和多模态数据的涌现，使得推荐系统在数据处理环节面临冗长、资源利用不合理以及传统数据处理框架难以满足多模态数据处理需求等问题。 课题挑战: 策略管理与优化：构建一套智能化系统，实现推荐策略的规范化定义、长期及离线评估、无效策略自动识别与下线，以及相关代码配置的下线。 自适应调优与故障诊断：针对推荐系统多样化业务负载，利用大模型能力完成系统及底层组件的参数和配置调优，并探索自适应故障诊断方案，提供全局视角的故障追踪、定位和分析能力。 成本与效率平衡：在推荐系统应用生成式技术时，解决模型训练和运行的高成本问题，平衡成本与效率，在有限资源下实现高效推荐。 全域数据处理：应对电商等横向全域场景下海量异构数据，提升和保障数据质量与准确性，标准化供给数据给全域推荐模型，并实现低成本跨端服务，同时，确保数据隐私与安全，合规使用数据。 数据存储与质量提升：研发低成本高性能存储引擎，设计灵活的Schema Evolution机制，实现数据高并发实时写入与训推一致性，深入探究数据质量与模型预测性能的量化关系，构建基于DCAI理念的数据和模型相关性分析工具及训练数据自动化处理链路。 多模态数据与异构计算：构建适用于推荐系统的多模态数据异构计算处理框架，解决数据读取、框架整合、高性能算子编排等问题，提高数据处理和模型训练效率，建立以Python为核心的开发者生态。 推荐大算力模型效率优化：随着大模型在CV/NLP/多模态以至于AGI领域的不断突破，推荐场景下的大算力驱动能够帮助模型更全面深刻理解用户偏好，进而更好地理解用户需求，挖掘用户潜在兴趣，进而带来更好地用户体验。更大规模的推荐模型需要更大的算力，如何平衡好算力开销和效果收益，需要架构和算法工程师深度Co-Design。

Team Introduction: Data AML is ByteDance's machine learning middle platform, providing training and inference systems for recommendation, advertising, CV (computer vision), speech, and NLP (natural language processing) across businesses such as Douyin, Toutiao, and Xigua Video. AML provides powerful machine learning computing capabilities to internal business units and conducts research on general and innovative algorithms to solve key business challenges. Additionally, through Volcano Engine, it delivers core machine learning and recommendation system capabilities to external enterprise clients. Beyond business applications, AML is also engaged in cutting-edge research in areas such as AI for Science and scientific computing. Research Project Introduction: Large-scale recommendation systems are being increasingly applied to short video, text community, image and other products, and the role of modal information in recommendation systems has become more prominent. ByteDance's practice has found that modal information can serve as a generalization feature to support business scenarios such as recommendation, and the research on end-to-end ultra-large-scale multimodal recommendation systems has enormous potential. It is expected to further explore directions such as multimodal cotraining, 7B/13B large-scale parameter models, and longer sequence end-to-end based on algorithm-engineering CoDesign. Engineering research directions include: Representation of multimodal samples Construction of high-performance multimodal inference engines based on the PyTorch framework Development of high-performance multimodal training frameworks Application of heterogeneous hardware in multimodal recommendation systems 1. Algorithmic research directions include: 2. Design of reasonable recommendation-advertising and multimodal cotraining architectures 3. Sparse Mixture of Experts (Sparse MOE) 4. Memory Network 5. Hybrid precision techniques 团队介绍： Data AML是字节跳动公司的机器学习中台，为抖音/今日头条/西瓜视频等业务提供推荐/广告/CV/语音/NLP的训练和推理系统。为公司内业务部门提供强大的机器学习算力，并在这些业务的问题上研究一些具有通用性和创新性的算法。同时，也通过火山引擎将一些机器学习/推荐系统的核心能力提供给外部企业客户。此外，AML还在AI for Science，科学计算等领域做一些前沿研究。 课题介绍： 大规模推荐系统正在越来越多的应用到短视频、文本社区、图像等产品上，模态信息在推荐系统中的作用也越来越大。 字节实践中发现模态信息能够很好的作为泛化特征支持推荐等业务场景，端到端的超大规模多模态推荐系统的研究具有非常大的想象空间。 期望在算法和工程CoDesign基础上，对多模态Cotrain、7B/13B大规模参数模型、更长序列端到端等方向进一步进行探索。 工程上研究方向包括多模态样本的表征、基于 pytorch 框架的高性能多模态推理引擎、高性能多模态训练框架的构建、异构硬件在多模态推荐系统上的应用；算法上的研究方向包括设计合理的推荐广告和多模态Cotrain结构、Sparse MOE、Memory Network、混合精度等。 1、负责机器学习系统架构的设计开发，以及系统性能调优； 2、负责解决系统高并发、高可靠性、高可扩展性等技术难关； 3、覆盖机器学习系统多个子方向领域的工作，包括：资源调度、任务编排、模型训练、模型推理、模型管理、数据集管理、工作流编排、ML for System等； 4、负责机器学习系统前瞻技术的调研和引入，比如：最新硬件架构、异构计算系统、GPU优化技术的引入落地； 5、研究基于机器学习方法，实现对集群/服务资源使用情况的分析和优化。

1.负责研发阶段工艺可靠性的研究及标准的制定； 2.结合PIE/UPD/CMOS/Design，准确诊断、定位以及解决各类工艺可靠性问题，并给出优化方向及措施； 3.总结/分析可靠性结果，结合器件参数退化趋势，准确建立及运用加速失效模型来预测产品的可靠性寿命； 4.同design一起，Control研发阶段Design for Reliability可靠性，提升相关电路的可靠性设计； 5.负责器件测试、数据分析与处理、编写测试报告、设计SOA评估及制定