Keysight Deploys AI to Automate Factory Production Scheduling

Keysight部署AI实现工厂生产排程自动化 Keysight Technologies利用RapidCanvas AI模型和LLM清洗车间历史数据并自动化运营叙述报告，从而提高了产量并节省了主管的日常工时。 面对Santa Clara新洁净室中关键的产品产能爬坡，加之依赖拥有25年历史的旧工单系统，Keysight曾饱受运营摩擦与混乱的电子表格数据困扰。通过将数据管道整合至Snowflake并与RapidCanvas合作，他们清洗了杂乱的交易数据，实现了精确到分钟级的生产排程。此外，他们还部署了根据主管写作风格训练的LLM，通过Microsoft Teams自动发送每日车间运营汇报，不仅每天早晨节省了数小时，还促进了员工的跨职能培训。 Keysight 在高难度新产品产能爬坡与圣克拉拉新厂房搬迁期间，因缺乏时间部署传统 MES，面临新员工不熟悉 25 年历史老旧追踪系统的严峻运营压力。 (0:04) 为消除手工电子表格与邮件沟通带来的低效摩擦，团队通过构建 ETL 管道与 Python 脚本，将分散在旧系统、SmartSheet 与 Excel 中的排班及认证数据整合至 SQL 数据库中。 (4:01) Keysight 携手 RapidCanvas 将排程引擎扩展并迁移至 Snowflake，通过 AI 自动清洗密集的工单“补录交易”数据，成功将排程精度从小时级细化至分钟级，并引入了实时查询交互式 Chatbot。 (5:40) 借助大语言模型（LLM）模拟车间主管风格自动生成每日运营任务报告，每天节省约 2 小时排程汇报撰写时间，将复杂的表格数据转化为一线工人清晰易懂的操作指令。 (8:27)

Faced with ramping up an intricate assembly in a new cleanroom without modern manufacturing execution software, Keysight Technologies integrated its legacy tracking systems with RapidCanvas AI to establish minute-level scheduling and automated natural language floor directives. Operational Pressures and Fragmented Legacy Infrastructure (0:04) Keysight encountered acute operational strain when launching its most complex assembly to date for a demanding customer requiring unprecedented production volumes. The ramp coincided with shifting operations to a new cleanroom facility in Santa Clara, California, and onboarding a technical workforce unfamiliar with Keysight's systems. Because time constraints prevented the deployment of a conventional Manufacturing Execution System (MES), operations relied on a 25-year-old job tracking system lacking a modern backend, supplemented by employee certifications in SmartSheet and shift schedules in Excel spreadsheets. With new technicians focused entirely on learning the intricate assembly process, transaction recording suffered. Frontline operations resorted to manual workarounds, email threads, and physically walking the floor to locate serialized units. The resulting lack of real-time visibility placed revenue and customer trust at risk, forcing leadership to seek an automated pipeline that could track units and estimate completion times without burdening technicians with antiquated software. Pipeline Integration and Automated Transaction Scrubbing (5:40) Keysight formed a cross-functional internal team—including an industrial engineer, a database administrator, a production scheduler, and a software engineer—to extract data from the legacy system into a local SQL Server using Python scripts and APIs. However, shop-floor supervisors initially bypassed the Python scheduler outputs because the dense mathematical matrices were difficult to interpret. Keysight subsequently partnered with RapidCanvas, migrating data into Snowflake to replicate and scale the scheduling engine. RapidCanvas's analytics resolved a fundamental data-quality issue: assemblers frequently logged multiple steps within seconds at the end of shifts to make up for missed entries. The platform automatically identified and scrubbed these "makeup transactions," enabling Keysight to calculate genuine cycle times. This automated data cleansing allowed the facility to transition from coarse hourly scheduling to minute-by-minute planning, supported by an interactive chatbot for real-time tracking queries. Automating Daily Floor Directives with LLMs (8:27) Even with refined schedule data, morning operations were bottlenecked by administrative overhead. Each morning, a shop-floor supervisor spent up to two hours reviewing unit statuses, rework needs, and engineering holds to draft a comprehensive text summary for Microsoft Teams. Because this narrative was the primary mechanism used to run floor operations, Keysight integrated a Large Language Model (LLM) trained on historical examples of the supervisor's writing style to draft the daily message automatically. The automated pipeline ingests raw schedule data from Snowflake and Excel, outputs a draft matching the supervisor's standard terminology, and allows her to perform a brief review before posting. According to Ethan Hunt, this system reclaimed hundreds of hours of supervisor time annually. Furthermore, translating complex spreadsheets into plain-language instructions improved operational clarity for technicians, many of whom spoke English as a second language. Technical Demarcation and Implementation Strategy (14:53) Addressing technical architecture during the presentation's Q&A, Hunt clarified the functional boundary between generative AI and machine learning. Keysight avoided using generative AI for numerical forecasting due to industry-wide accuracy limitations; instead, traditional machine learning models and deterministic algorithms calculate predictive schedules and identify cross-training bottlenecks, while LLMs handle narrative generation and code creation prompts. Hunt offered four core recommendations for manufacturing leaders deploying AI: centralize raw data into cloud warehouses such as Snowflake early to accelerate partner onboarding; engage AI partners promptly to validate architecture; maintain active change management with frontline workers who may not speak up when tools fail; and treat AI as an evolving operational baseline rather than a static fix. Keysight's deployment demonstrates that manufacturing AI projects succeed when advanced data cleaning is paired with intuitive, frontline-accessible communication. By using machine learning for quantitative cycle-time prediction and generative AI strictly for narrative synthesis, the team stabilized an urgent production ramp, though ongoing operational performance remains tied to continuous user feedback and adoption.

是德科技（Keysight Technologies）在面临复杂新品加速量产、生产线迁至加州圣克拉拉新洁净室以及大量新员工入职的压力下，由于缺乏现代制造执行系统（MES），陷入了数据失真与调度混乱的困境。主讲人伊森·亨特（Ethan Hunt）阐明，团队通过构建集成数据管道并联合RapidCanvas引入AI，不仅清洗了制造端操作误差以实现分钟级精确排程，更利用大语言模型（LLM）生成现场自然语言工单，成功将算法调度转化为一线工人高效执行的日常生产力。 新产线扩张压力与遗留系统的透明度断层 (0:04) 在新产品设计尚未完全定型且装配复杂度创下新高的情况下，是德科技需要满足苛刻客户远超以往的订单需求。与此同时，制造产线整体迁入圣克拉拉的新洁净室，大量缺乏企业历史经验的新组装工人和技术人员加入团队。管理层不希望新员工把精力消耗在摸索繁复过时的老旧软件上，而是专注于学习复杂的装配工艺。然而，生产现场长期依赖一套已有25年以上历史的过时工单交易系统，该系统既没有现代后端数据库，也缺乏完整的制造执行系统（MES）支持。 在这种条件下，运营团队最初只能依靠邮件链、跨部门会议和分散的电子表格等人工手段推进工作。由于新员工在培训期内产生的数据录入质量低下，人工追踪机制迅速失效，导致产线工件位置不明、完工时间难以预测、客户信任度受损，现场管理人员甚至开始彻底忽视系统生成的低质量数据。这促使是德科技确立了明确的技术目标：打通遗留系统数据、清洗错误记录并重建透明的排程机制。 数据整合与补录行为清洗实现精细排程 (4:01) 为了打通信息孤岛，是德科技组建了一支由项目经理、排程员、工业工程师、软件工程师及数据库管理员组成的跨职能团队。技术人员通过API接口与Python脚本，将SmartSheet中的技术认证信息、Excel中的班次和休假记录，以及遗留系统中的工单交易流水统一提取至本地SQL Server，随后整体迁移至Snowflake云数据仓库。初期基于Python构建的前向排程算法虽能推算工件进展，但现场主管反馈其输出过于晦涩，难以真正指导车间作业。 随后，Keysight引入RapidCanvas平台复制并扩展了排程逻辑，重点解决了数据质量痼疾。平台的数据科学功能敏锐识别出一线工人常在交接班时集中“补录”丢失步骤的行为——装配人员为了弥补先前遗漏，往往在数秒内连续点击提交多道工序。RapidCanvas自动识别并剔除了这些失真的伪交易记录，从而还原了真实的工件生产节拍（cycle time）。这一清洗使车间排程从粗放的小时级粗粒度跃升为精确的分钟级细粒度，同时引入了允许员工直接用自然语言查询工件位置与工时排班的交互式助手。 引入大语言模型消除现场自然语言执行壁垒 (8:27) 算法模型的高精度并不等同于现场的有效执行。在圣克拉拉工厂的实地调研中，团队发现车间主管每天必须耗费长达2小时，从密集的排程表格中排查返工优先级与工程暂停项，再手写成长篇摘要发布在Microsoft Teams中作为调度依据。对于许多以英语为第二语言（ESL）的组装人员而言，密集的算法表格宛如天书，只有主管亲笔撰写的自然语言文本才是真正驱动生产的指令。 为此，团队利用主管历史撰写的车间日报训练LLM，使其深度掌握主管的表达用词、格式与专业语境。自动化流水线直接从底层排程数据生成车间运营叙事草稿，主管清晨只需花数分钟审阅微调即可分发。此举每年为主管挽回数百小时繁琐的人工报表编写时间，彻底消除了表格数据与一线理解之间的认知落差。同时，平台对多技能交叉培训需求的智能分析，协助管理层精准消除关键技术瓶颈岗位的人员依赖，有效削减了周末加班负荷。 预测性机器学习与生成式文本的架构分工 (12:21) 伊森·亨特在经验分享与问答环节中着重阐释了制造车间部署AI的核心技术架构与实施策略。针对行业内对“生成式AI直接预测数值易产生幻觉”的疑虑，亨特明确澄清了Keysight的架构分工：系统底层的数值建模、生产节拍预测与关键路径优化，完全由传统机器学习及算法引擎承担；LLM的职责则严格界定在将分析结果转化为自然语言报告、交互式查询以及辅助编写分析代码等交互层面。 基于本次转型实践，亨特为制造企业提出了四项落地准则：一是及早将车间原始数据集中到Snowflake等云端，避免在本地中间库上延误工期；二是及早引入外部AI生态伙伴评估数据成熟度，加速原型构建；三是强化现场变革管理，因一线技术人员受制于语言障碍或技术畏难情绪往往不会主动反馈系统痛点，必须持续跟进使用体验；四是将AI视为不断提升运营基准的能力底座，而非一劳永逸的短期修补工具。 主讲人伊森·亨特强调，借助AI自动化决策能够大幅压缩制造业现代化转型的周期，但该项目的成功高度依赖于对数据噪音的清洗以及对一线操作习惯的尊重。尽管LLM显著降低了调度信息的使用门槛，但对于更为底层、多变的突发设备故障或供应链动态扰动，如何让传统排程算法与生成式交互实现更深度的双向闭环联动，仍是此类工业AI架构有待进一步探索的课题。

## Background 该内容由 Ai4 频道于 2025 年 8 月 28 日发布在 YouTube 平台上。视频记录了是德科技（Keysight Technologies）数字化转型经理 Ethan Hunt 的主题分享，介绍了 Keysight 如何在不推倒重来 25 年老旧底层系统的前提下，携手 RapidCanvas 借助 AI 决策、自动化数据清洗与大语言模型重塑生产排程并释放一线产能。 ## TL;DR 面对古董级工单系统、产线搬迁与新手员工技能不足带来的产能爬坡危机，Keysight 团队自研的算法排程最初因过于复杂而遭遇一线主管弃用。随后团队携手 RapidCanvas 将数据迁移上云，通过算法过滤批量补录的异常工时，将调度精度跃升至分钟级，并创新性地利用大语言模型微调自动生成主管口吻的每日排产叙述。整个转型不仅每天为主管节省数小时晨间排查与文书工作，还通过精准技能分析削减了加班，探索出一条传统制造业低门槛落地 AI 的敏捷路径。 ## Atmosphere & style 演讲整体呈现出极其接地气、坦诚且充满实战智慧的工业科技叙事氛围。Ethan Hunt 以幽默且不避讳失败的口吻，生动复盘了一线主管直言“我们根本不用它”的尴尬挫败，以及随后在车间一线迎来利用 LLM 自动生成排产叙述的“尤里卡时刻”。整场分享既有应对严苛交付压力的紧迫感，又有拆解技术与人际变革冲突的理性洞见，给观众带来扎实、真诚且充满启发性的工业 AI 落地体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=4s) Keysight 半导体基础设施解决方案部门在 25 年老旧工单系统、无后端数据库、产线搬迁及大量新员工涌入的复杂环境下，面临产能紧急爬坡的严峻挑战。 - [03:03](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=183s) 数字化转型团队确立了四大核心目标：清洗数据推导真实工序节拍、建立人员与工位技能基准、输出清晰简明的周度排程，以及引入 AI 实现调度全流程自动化。 - [04:01](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=241s) 初期内部自研的 Python 正向排程算法因生成的复杂表格脱离车间一线习惯，导致主管仍依赖纸质交接与经验排产，暴露出技术与现场脱节的瓶颈。 - [06:31](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=391s) 联合 RapidCanvas 迁移至 Snowflake 云端后，算法成功识别并过滤了下班前集中“批量补录”的异常打卡数据，推动排程精度从小时级跨越到分钟级精细化调度。 - [07:40](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=460s) 基于调度数据上线自然语言交互 Chatbot，使管理人员能够直接通过对话查询工单位置、排班情况并即时生成分析图表。 - [07:40](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=460s) 利用微调后的大语言模型（LLM）将复杂的底层排程数据自动转化为符合主管行文风格的每日生产叙述（Daily Narrative），每天为主管节省 2 至 3 小时的高强度文书工作。 - [10:35](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=635s) 借助平台对工位瓶颈与技术员资质重合度的深入分析，团队精准定位交叉培训需求，有效打破了少数关键技术人员必须高强度周末加班的恶性循环。 - [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 制造业落地 AI 的首要经验是尽早将多源数据集中并迁移至云端数据仓库（如 Snowflake），这是快速对接 AI 平台并实现分钟级模型构建的前提。 - [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 及早引入专业外部 AI 合作伙伴有助于在架构设计和上云路径上少走弯路，最大化把握业务落地时机（Speed to Opportunity）。 - [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 车间一线变革管理必须深入现场，避免使用晦涩的技术术语，通过高频且真实的沟通打破语言与心理壁垒，获取真实的使用反馈。 - [14:38](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=878s) 在架构设计上明确区分生成式 AI 与传统分析的分工：LLM 负责自然语言叙事与辅助写码，底层严谨的数值分析与精准排程则交由确定性模型，避免单纯依赖生成式 AI 预测数值带来的幻觉风险。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=4s) 复杂制造困境与数字化破局 这场转型的开端面临着极其苛刻的现实：Keysight（是德科技）半导体基础设施解决方案部门接到了一位高要求客户的全新产品订单，其装配复杂度与产能需求远超以往任何项目。然而，车间依赖的是一套有着 25 年以上历史、甚至连后端数据库都没有的老旧工单事务系统；与此同时，产线引入了大批新员工，他们缺乏在关键节点准时记录物料移动的操作纪律。面对“极限交付、古董系统、新手团队”的三重挑战，团队必须探索如何利用 AI 将这些破碎的环节整合为“大于各部分之和”的整体。 在加入该部门时，团队正处于为关键客户紧急爬坡产能的关键期。雪上加霜的是，整个生产爬坡是在产品设计尚未完全定型、且产线正搬迁至加州圣克拉拉主园区几英里外全新无尘车间的多重混乱状态下进行的。由于大规模扩张，新招募的装配工和技术人员大多此前从未在 Keysight 工作过。团队认为，绝不能让新手把宝贵的时间浪费在摸索过时的跟踪系统和繁杂的电子表格上，他们的精力必须 100% 集中在掌握产品组装技能上。这导致数据录入和工单事务质量天然存在缺陷。 依靠开不完的会议、往来邮件和混乱的 Excel 表格，生产勉强维持，但客户的挫败感与日俱增，运营团队也逐渐对系统产生的数据视而不见，公司的周度出货量与营收面临严重威胁。为了消除噪音、重塑客户信心并稳定每周产出，数字化团队设立了核心使命：彻底拉通数据，提供实时的产线透明度，明确掌握“产品当前在哪个工位、还要多久、何时能完工交付”。 ```mermaid flowchart TD accTitle: Keysight 生产调度架构演进 accDescription: 展示从多源数据整合到 AI 驱动排程与大模型叙述生成的端到端架构 subgraph DataSources["多源底层数据采集"] A["25年历史老旧工单系统<br/>(无后端数据库/Python采集)"] B["Smartsheet<br/>(员工技能与培训资质API)"] C["Excel 杂项数据<br/>(工时/班次/请假/节拍)"] end subgraph DataPlatform["数据集成与清洗层"] D["本地 SQL Server / Snowflake"] E["RapidCanvas 数据清洗引擎<br/>(剔除末端批量补录秒级异常数据)"] end subgraph AIScheduling["AI 决策与应用层"] F["分钟级排程引擎<br/>(Python 正向排程算法)"] G["自然语言问答 Chatbot<br/>(工单查询/报表即时生成)"] H["大语言模型 (LLM)<br/>(自动生成主管风格排产叙述)"] end A --> D B --> D C --> D D --> E E --> F F --> G F --> H ``` --- ### [03:03](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=183s) 明确四大核心任务目标 为了达成使命，团队梳理并确立了必须逐一攻克的四个具体目标： - **抓取并清洗历史与实时事务数据**：从毫无数据库支持的老旧系统中提取数据，并通过算法剔除质量低劣的操作记录，从而推导出具有实际指导意义的真实生产周期（Cycle Time）。 - **建立人员与工位能力基准数据**：系统化录入员工技能熟练度、已获认证的工位资质、各工序标准工时，以及涵盖白班、中班、夜班与休假计划的排班表。 - **输出可执行的周度生产排程**：以极简明了为原则，实时回答三大核心问题：“产品目前在哪里？”、“在当前工位还需停留多久？”以及“整机预计何时完工？”。 - **引入 AI 实现全流程自动化与智能分析**：利用人工智能自动化处理调度数据流程，并生成深入的运营效能洞察与分析。 --- ### [04:01](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=241s) 内部自研排程的瓶颈与一线脱节 为推进项目，部门迅速集结了一支跨职能专家团队：项目经理负责整体推进；新招募的生产调度员先以手工排程支撑业务并梳理业务逻辑；软件工程师编写 Python 脚本基于当前在制品状态进行正向排程迭代；工业工程师负责现场工时测算与产线动态分析；经验丰富的数据库管理员（DBA）则利用 Python 脚本从无数据库的老旧系统中抓取数据并导入 SQL Server。 团队通过 API 将存储在 Smartsheet 上的技术员培训资质打通，并通过基础 ETL 流程将记录在 Excel 中的班次、工位标准工时和员工休假数据汇集到 SQL Server。Python 脚本基于这些数据顺利输出了在制品下一步流向与预计完工时间。技术团队本以为大功告成，但随后在一次管理层会议上遭遇了当头棒喝。 当负责人兴致勃勃地汇报项目进展和海量数据时，运营主管冷淡地回应："Yeah, but you know, we're not really using it."（“是的，但我们其实根本没在用这些数据。”）团队调查发现，算法生成的复杂表格极难被一线理解，主管们依然依赖纸质交接日志和 Excel，每天亲自在产线上来回走动寻找具体序列号的机器，而工人的派工指令全凭车间主管的“脑中记忆”和个人经验。 --- ### [06:31](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=391s) 携手 RapidCanvas 实现分钟级排程数据清洗 意识到系统与现场脱节后，Keysight 决定引入 AI 数据平台合作伙伴 RapidCanvas 协助破局。团队将全部数据迁移至 Snowflake 云端，RapidCanvas 团队在短短几天内就将原有的本地 Python 算法无缝接入平台并成功复现。 随后，RapidCanvas 的数据科学家接过了技术主导权，并迅速攻克了数据质量顽疾。通过对底层打卡记录的算法识别，系统成功捕获并剔除了员工在下班前几秒钟内连续“批量补录”的大量虚假步骤数据。利用清洗后的高质量数据，算法重新推导出了高精度的实际工序节拍。排程精度由此从原先粗糙的“按小时排产”直接跨越到了“按分钟级精细化调度”，排程算法输出的准确度迎来了质的飞跃。 - **快速迁移验证**：依托 Snowflake 数据源，数天内完成本地脚本向云端 AI 平台的迁移与复现。 - **异常事务清洗**：自动识别并过滤数十秒内连续触发的补录数据，还原本写真实工时。 - **调度颗粒度跃升**：工位排程由小时级提升至分钟级，显著降低了工单积压与等待浪费。 --- ### [07:40](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=460s) 从自然语言问答到大模型日报自动化 在完成算法优化的基础上，团队进一步探索生成式 AI 的前沿应用。他们首先开发了基于调度数据的自然语言交互助手，产线管理人员和调度员不再需要互相频繁追问，而是直接向 Chatbot 提问（例如：“X 序列号的产品目前在哪？”、“Jane Doe 下周排班多少小时？”、“Y 产品预计何时下线？”），甚至可以通过提示词指令直接实时生成分析图表。 在圣克拉拉现场的一次研讨中，调度员提出了一个痛点：车间主管每天清晨都需要花费 2 到 3 个小时排查在制品位置、识别需要工程介入的返修机台、整理各工位优先级，最后在 Microsoft Teams 中手写发布一段长篇排产叙述（Daily Narrative）来指挥生产。团队随即迎来“尤里卡时刻”——决定利用大语言模型（LLM）彻底将这一耗时流程自动化。 团队提取了主管以往编写的高质量叙述范本，将主管特有的行文语气、专业术语和排版习惯喂给 LLM 进行微调训练，并结合运营团队的需求反复调试输出格式与触发时间。系统最终实现了每天清晨自动提取分钟级排程数据并生成叙述草稿，主管只需花几分钟快速审阅微调即可一键发布至技术员群组。过去如天书般的冷冰冰 Excel 数据，被转化为了产线员工一目了然的“行动指南”，每天为主管直接节省数小时的高强度文书工作。 ``` [原始调度数据: 复杂 Excel 矩阵/时间戳] │ ▼ (LLM 自动解析与风格化转换) [各工位执行指令: 包含机器优先级、返修件指引及工程支援标记的自然语言叙事] ``` --- ### [10:35](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=635s) 业务成效与关键价值落地 该项目的推进带来了超出初期预期的硬性与软性综合收益： - **排程精度与可读性双重突破**：不仅消除了工时虚标，而且将复杂算法以全员秒懂的自然语言形式下发，消除了技术落地壁垒。 - **技能矩阵洞察与加班削减**：通过 RapidCanvas 分析各工位瓶颈与技术员资质重合度，精准定位急需开展交叉培训的工位，打破了少数关键技术员因技能垄断而必须周末高强度加班的恶性循环。 - **释放管理层生产力**：不仅每天为主管节约 2-3 小时的晨间文书时间，全年节省数百小时工时，更通过 Chatbot 自助服务免去了大量日常打扰。 - **组织敏捷性建立**：建立了基于现代 AI 平台的敏捷开发范式，为后续扩展到更多制造运营场景打下了坚实的技术与流程基础。 --- ### [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 制造业落地 AI 的实战建议 回顾整个数字化转型历程，负责人为面临类似场景的企业分享了四条核心避坑与实践心得： - **数据集中与尽早上云**：不要在本地临时服务器上停留太久。如果重新来过，团队会选择第一天就直接将多源数据汇集至 Snowflake 等云数据仓库，这是快速对接 AI 平台并实现分钟级模型构建的前提。 - **及早引入专业 AI 合作伙伴**：越早引入合作伙伴越好，即使只是进行前期现状评估。优秀伙伴可以在企业花冤枉钱之前就架构设计和上云路径提供关键指导，抢占业务机遇（Speed to Opportunity）。 - **重视一线变革管理与高频沟通**：制造车间中很多工人的母语并非英语，过高深的技术术语容易引发一线抵触甚至隐瞒系统缺陷。必须深入现场高频询问真实使用体验，“问得越勤，得到的真话就越多”。 - **将 AI 视为提升能力下限的长期杠杆**：即便团队内部没有专属的数据科学家，也可以通过低门槛 AI 平台赋能业务人员用自然语言生成代码。不要将 AI 视作一蹴而就的“速效药”，而应将其作为持续拔高组织基础能力下限（Lift the floor of capabilities）的战略工具。 ### [14:38](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=878s) 总结与技术架构答疑：大语言模型与传统分析的分工 在演讲的结尾，Ethan Hunt 总结道，通过引入现代 AI 工具与自动化工作流，企业不仅能够探索并实现全新的业务目标，而且在推进速度上远快于传统的固守陈旧系统的方式。在向听众展示 RapidCanvas 进一步了解信息的二维码时，他还幽默地提醒大家“动作要快”，随后正式进入现场观众问答环节。 在问答中，一位现场观众提出了关于技术架构的疑问，询问 Keysight 在方案中究竟使用的是生成式 AI（GenAI）还是传统的机器学习（Machine Learning）模型。该观众补充说明了自己团队遇到的痛点：在尝试使用生成式 AI 预测具体实际数值时，经常会遇到准确率不足、产生幻觉的问题。Ethan 对此坦承，底层最深入的架构实现建议直接前往 RapidCanvas 的展位向技术专家深入咨询，但他结合实际落地场景阐明了 LLM（大语言模型）的核心定位。 ```mermaid flowchart TD accTitle: Keysight 生产调度方案中 LLM 与数据分析的分工 subgraph RapidCanvasPlatform["RapidCanvas 平台技术应用"] A["生产排程与底层数值运算"] --> B["自动化代码生成与数据分析<br/>（Prompt 辅助开发）"] A --> C["每日生产运营总结<br/>（LLM 生成 Daily Narrative）"] end B --> D["生产现场决策与排程下发"] C --> D ``` 根据 Ethan 的说明，系统在不同层面进行了明确的技术分工： - **生成日常运营叙述（Daily Narrative）**：系统利用大语言模型（LLM）将复杂的排程变动和工厂状态转化为自然语言报告，直接为产线主管生成清晰易读的每日晨会摘要。 - **辅助生成分析代码**：利用 LLM 的对话式 Prompt（提示词）功能来协助编写数据分析与调度逻辑的代码。 - **严谨数值分析与模型计算**：底层的确定性数据分析与机器学习模型处理精准的生产预测与排程约束，避免了单纯依赖生成式模型预测数值时准确度不佳的缺陷。 --- ### [15:51](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=951s) 是德科技（Keysight）的业务体量与全球制造布局 随后，现场观众就 Keysight 的企业规模与业务背景进行了提问。Ethan 向参会者介绍了 Keysight Technologies（是德科技）的整体体量与制造版图： - **公司营收体量**：Keysight 是一家年营收约 50 亿美元的行业领军企业。 - **核心业务领域**：公司主要专注于电子测试与测量（Electronic Test and Measurement）设备的设计与制造；尽管 Ethan 负责支持的特定制造厂区有其特殊的专业制造任务，但这构成了公司最核心的业务基石。 - **全球制造足迹**：Keysight 在全球范围内拥有广泛的布局，制造和运营分支遍布美国本土、欧洲以及马来西亚等多个国家和地区。 ## Notable quotes > [04:01](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=241s) Yeah, but you know, we're not really using it. — 车间运营主管 > [00:04](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=4s) 绝不能让新手把宝贵的时间浪费在摸索过时的跟踪系统和繁杂的电子表格上，他们的精力必须 100% 集中在掌握产品组装技能上。 — Ethan Hunt > [00:04](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=4s) 彻底拉通数据，提供实时的产线透明度，明确掌握‘产品当前在哪个工位、还要多久、何时能完工交付’。 — Ethan Hunt > [03:03](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=183s) 从毫无数据库支持的老旧系统中提取数据，并通过算法剔除质量低劣的操作记录，从而推导出具有实际指导意义的真实生产周期。 — Ethan Hunt > [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 深入现场高频询问真实使用体验，‘问得越勤，得到的真话就越多’。 — Ethan Hunt > [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 不要在本地临时服务器上停留太久。如果重新来过，团队会选择第一天就直接将多源数据汇集至 Snowflake 等云数据仓库。 — Ethan Hunt > [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 优秀伙伴可以在企业花冤枉钱之前就架构设计和上云路径提供关键指导，抢占业务机遇（Speed to Opportunity）。 — Ethan Hunt > [12:21](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=741s) 不要将 AI 视作一蹴而就的‘速效药’，而应将其作为持续拔高组织基础能力下限（Lift the floor of capabilities）的战略工具。 — Ethan Hunt > [14:38](https://www.youtube.com/watch?v=rBMWKxy7zK4&t=878s) 通过引入现代 AI 工具与自动化工作流，企业不仅能够探索并实现全新的业务目标，而且在推进速度上远快于传统的固守陈旧系统的方式。 — Ethan Hunt ## Mentioned Ethan Hunt, Keysight Technologies (是德科技), RapidCanvas, Ai4, Snowflake, SQL Server, Python, Smartsheet, Microsoft Teams, Microsoft Excel ## Video description Ethan Hunt, Manager - Digital Transformation, Keysight Technologies Abstracts How can manufacturers scale output, improve quality, and modernize operations—without ripping out legacy systems or waiting for perfect data? In this session, I’ll share how we at Keysight partnered with RapidCanvas to transform our production scheduling using AI-powered decisioning, automated data scrubbing, and large language models. The talk will walk through the real-world impact—boosted output, reduced waste, and how supervisors got their mornings back. You’ll learn how to apply AI starting from your current systems, no overhaul required. If you’re ready to make AI real (and useful) on the factory floor, this one’s for you. [Source](https://www.youtube.com/watch?v=rBMWKxy7zK4)

[4:01] 车间运营主管: Yeah, but you know, we're not really using it. [0:04] Ethan Hunt: 绝不能让新手把宝贵的时间浪费在摸索过时的跟踪系统和繁杂的电子表格上，他们的精力必须 100% 集中在掌握产品组装技能上。 [0:04] Ethan Hunt: 彻底拉通数据，提供实时的产线透明度，明确掌握‘产品当前在哪个工位、还要多久、何时能完工交付’。 [3:03] Ethan Hunt: 从毫无数据库支持的老旧系统中提取数据，并通过算法剔除质量低劣的操作记录，从而推导出具有实际指导意义的真实生产周期。 [12:21] Ethan Hunt: 深入现场高频询问真实使用体验，‘问得越勤，得到的真话就越多’。 [12:21] Ethan Hunt: 不要在本地临时服务器上停留太久。如果重新来过，团队会选择第一天就直接将多源数据汇集至 Snowflake 等云数据仓库。 [12:21] Ethan Hunt: 优秀伙伴可以在企业花冤枉钱之前就架构设计和上云路径提供关键指导，抢占业务机遇（Speed to Opportunity）。 [12:21] Ethan Hunt: 不要将 AI 视作一蹴而就的‘速效药’，而应将其作为持续拔高组织基础能力下限（Lift the floor of capabilities）的战略工具。 [14:38] Ethan Hunt: 通过引入现代 AI 工具与自动化工作流，企业不仅能够探索并实现全新的业务目标，而且在推进速度上远快于传统的固守陈旧系统的方式。

[5:40] Operations Manager: Yeah, but you know, we're not really using it. [5:40] Where is Unit X? [5:40] How many hours is Jane scheduled for next week? [8:27] Station 4: Prioritize Unit 10482 for rework before moving to final calibration... [14:38] Ethan Hunt: might want to hurry [14:53] Ethan Hunt: a lot faster than if you don't.

[0:04] A new product for a demanding customer [0:06] with complexity and volume beyond [0:08] anything we had done before. A 25 plus [0:11] year old aging job transaction system [0:14] for tracking builds to our production [0:15] line. And a new workforce lacking the [0:19] training and operational discipline to [0:20] perform the move transactions when we [0:22] really need them to. How could we use AI [0:25] to put these pieces together to create [0:27] something that's greater than the sum of [0:28] the parts? [0:30] Uh, show of hands. How many here work in [0:31] manufacturing in some way? Oh, nice. [0:34] Okay. I actually didn't expect that many [0:35] of you to be here. Great. Thank you for [0:36] coming. All right. So, today we'll go [0:38] over the circumstances leading up to our [0:40] mission, share our goals, and lay out [0:42] the mission objectives. We'll give you a [0:44] glimpse into how we pulled together our [0:46] data and partnered with Rapid Canvas to [0:48] create something that was greater than [0:49] the sum of the parts. And finally, I'll [0:52] share some of the results we achieved [0:53] and a few tips for when you find [0:55] yourself executing a similar mission. [0:59] When I joined the semiconductor [1:01] infrastructure solutions group, we were [1:03] in the throws of ramping a product [1:04] critical to one of our customers. It's [1:07] the most complex assembly we've ever [1:08] attempted and the customer was very [1:10] demanding and they needed volume beyond [1:12] anything we were really, you know, [1:13] accustomed to. So, we didn't really have [1:16] time to implement traditional means of [1:18] manufacturing, you know, execution [1:20] systems to track the builds, uh, [1:22] advanced production scheduling, things [1:24] like that. So, we decided to pull [1:26] together just enough information to keep [1:28] the product moving through the line, [1:30] avoid major disruptions and quality [1:32] issues, and try to stabilize the [1:34] operation. The crazy thing though was [1:36] this was all in the middle of of also [1:38] finalizing the product design. It really [1:39] wasn't mature yet, and moving production [1:42] to a new clean room facility a few miles [1:44] away from our main campus in uh Santa [1:46] Clara, California. So, um because of [1:49] this expansion, uh it really meant [1:51] hiring new assemblers and technicians. [1:53] Most of them had never worked at Ksite [1:55] before and this meant that you know we [1:58] had to do do a lot of training with them [2:00] and it didn't make sense to spend a lot [2:01] of their time learning how to navigate [2:03] spreadsheets and antiquated tracking [2:04] systems. Their focus really needed to be [2:07] on learning how to do the product [2:08] builds. [2:10] So um navigating countless spreadsheets [2:13] just didn't make sense. So we knew [2:14] transaction and data quality from the [2:16] team would be suboptimal. [2:18] So, we were squeaking by uh overall, but [2:21] the countless meetings, the emails, uh [2:23] the the spreadsheets were not really [2:25] getting the job done reliably. The [2:27] customer was getting frustrated and [2:29] operations was ignoring a lot of the [2:31] data we were actually producing for [2:32] them. Uh our revenue was at risk. We [2:35] weren't sure we'd be able to get the [2:36] right volumes out on a weekly basis. So, [2:39] the team's mission was to pull together [2:41] the information needed to give us [2:42] visibility of where the units are in the [2:44] production line and when they would be [2:46] done. [2:48] If we could cut down the noise, the [2:50] meetings, the spreadsheets, the emails, [2:52] uh we could regain the cusp of the [2:53] customer and stabilize our weekly [2:55] output. To make this happen, we had [2:57] specific objectives we needed to meet to [2:59] declare mission accomplished. [3:03] So, first we needed to secure the [3:05] necessary transaction data from our [3:07] antiquated system, which has no database [3:09] on the back end, believe it or not, and [3:12] scrub that poor transaction data we [3:13] talked about out of out of the uh the [3:16] data flows uh in order to get a more [3:19] meaningful cycle time. [3:21] Objective two, we needed to create input [3:24] data for worker training proficiency, [3:26] which workstations they had been trained [3:28] on. Uh, and also the expected cycle [3:30] times for the products and the, uh, [3:32] worker schedules, when would they take [3:34] time off, uh, who was working day day [3:36] shift, swing shift, night shift, all of [3:38] those things. [3:40] Third, we h we wanted to produce an [3:42] executable weekly schedule to ensure [3:44] we're on track to meet our output goals. [3:46] At this point, we're trying to keep it [3:47] very simple. Where's the product? How [3:48] much longer at the workstation? when's [3:50] it going to be done? And finally, [3:54] we needed to leverage the power of AI to [3:55] automate the scheduler data process to [3:58] generate some analytics of our [3:59] performance. [4:01] Fortunately, we had a team of subject [4:04] matter experts involved that could piece [4:06] together information about uh this an [4:08] information solution that would support [4:09] our ramp goals. So, we had a project [4:12] manager who could bring the team [4:13] together to try to execute the mission. [4:16] We had a production scheduler uh that I [4:18] hired and brought them in to run the [4:21] schedule manually for a while, but also [4:22] to help us build something greater. [4:25] We had a software engineer working on [4:27] Python scripts to produce a viable [4:29] schedule. So, they were taking a look at [4:31] all that data I just talked about and [4:32] trying to generate, you know, iterate [4:34] through and generate where the units [4:35] would be and where they should go. And [4:37] then we had an industrial engineer that [4:39] had already spent a lot of time uh [4:41] watching some of the early builds to [4:43] understand cycle times and some of the [4:44] dynamics in the production area. And [4:47] then finally when I joined the [4:48] organization I brought a DBA I've been [4:50] working with uh for a long time with me [4:52] and he was able to use Python to harvest [4:54] that transaction data out of that [4:56] antiquated system and also put put that [4:58] into the uh SQL server. So the small [5:02] team of people were able to pull [5:03] together the elements needed to produce [5:05] a production schedule. Technician [5:07] training logs were in Smart Sheet, which [5:08] if any of you have used it, it's really [5:10] more of like a a project management type [5:12] software, but we found an API and we're [5:14] able to pull that data into the SQL [5:16] server. And the expected cycle times, [5:18] shift information, workstation [5:20] information, worker time off. Like I [5:22] said, all that stuff was was managed in [5:24] Excel. And we just had a a real basic [5:26] ETL to pull those things into the SQL [5:28] server. So, so with all the data in [5:32] place, the Python scripts could be used [5:33] to look at the current state of the [5:35] production jobs and then forward [5:36] schedule where the unit should go next [5:38] and estimate job completion dates. So, [5:40] we thought the output looked great, the [5:42] data made sense, but something was still [5:45] missing. And I remember I was at a staff [5:47] meeting uh with the ops manager, my [5:50] manager, and you know, the rest of the [5:52] staff. And I'm I'm going on about how [5:54] great this project is going now. we've [5:56] got all this data and and the ops [5:57] manager says, "Yeah, but you know, we're [5:58] not really using it." And I thought, [6:00] "Okay, that's that's a problem." So, uh, [6:03] as it turned out, um, the the data we [6:05] were producing was really difficult for [6:07] them to digest. And they were still [6:09] using Excel files and turnover logs to [6:11] try to figure out where the units were. [6:13] They were still manually walking around [6:15] the production line to look individually [6:16] where a serialized unit was. And then [6:19] most of the worker directives were still [6:20] coming you know directly from the minds [6:22] of the of the actual u production sup [6:25] super supervisors. So we thought okay we [6:27] need to take another look at this and [6:28] this is where we thought rapid canvas [6:30] might be able to help us uh some more. [6:32] So we brought them back in to see if [6:34] their data scientists and their platform [6:36] could help us further automate the [6:37] execution of theuler code uh and also [6:40] help us introduce an array of analytics [6:42] to see how we were doing. So, um, and [6:45] this is where it gets fun. Within days, [6:47] uh, Rapid Canvas, uh, they were able to [6:49] actually import all the Python code we [6:51] had already created, connected to our [6:53] data source in Snowflake after we we [6:55] pushed all of our data up there, and [6:57] they could replicate the results of our [6:58] locally run Python scripts um, uh, [7:01] pretty pretty quickly in their platform. [7:04] Um, they were able to identify erroneous [7:06] data where the workers were executing [7:08] transactions only seconds apart. And [7:10] when we knew these were wrong and they [7:11] were probably makeup transactions uh [7:14] that they were putting in at the end of [7:15] the build, we could use that scrub [7:18] transaction data to generate more [7:20] accurate cycle times. And this would [7:21] help obviously with generating more [7:23] accurate schedule. We could now schedule [7:25] the workstations by the minute instead [7:27] of hourly. So not only did Rapid Canvas, [7:29] you know, help us import the code, their [7:31] data scientist kind of took over and we [7:33] said, "Here's the keys." And he started [7:35] improving things almost immediately. [7:40] Even with that, I still felt like, okay, [7:42] I want more AI in the AI, like I wanted [7:44] something that was a little bit more [7:45] compelling to show to my management. And [7:47] so, one of the funner things that we [7:49] were able to accomplish is we, uh, now [7:52] have the ability to ask questions of [7:53] theuler data. So, uh, you know, where is [7:56] unit X or how many hours is Jane do [7:59] worker going to be working next week? [8:01] Uh, you know, when will unit Y be [8:03] finished? [8:04] people could actually now consult the [8:06] chatbot instead of peppering, you know, [8:08] our scheduler, the human scheduler with [8:11] questions about what was happening. They [8:13] can create analytics and charts directly [8:14] through the chat prompt. So, um, we I [8:18] this is where when I really felt like we [8:20] were using AI. The truth is we'd be been [8:22] using it all along to to do a lot of the [8:24] coding, but this was this was really [8:25] fun. So once we'd proven the capability [8:28] of the platform, we decided to keep [8:30] going and I remember we rapid came back [8:33] out for some more face-to-face sessions [8:35] uh at our Santa Clara site and we we [8:37] were at a point now where we were saying [8:38] okay how can we further optimize this [8:40] and we were going around the table with [8:41] different ideas and the scheduler says [8:43] hey you know uh every day uh the [8:46] supervisor uh she takes a look at where [8:48] units are and and she spends a couple [8:51] hours writing up like a daily narrative [8:53] and this narrative had priorities for [8:55] which workstations you know, needed to [8:57] be attended to where we had rework units [8:59] where we really needed engineering to [9:01] evaluate them and and then she would [9:03] take this this, you know, long narrative [9:05] that took a couple hours to create and [9:06] go and post it in teams. And this is how [9:07] we're, you know, running the [9:08] manufacturing line. And so we we all had [9:11] kind of a Eureka moment and we said, [9:12] "Hey, you know, we could probably use an [9:14] LLM to take the schedule data, not only [9:17] make a more accurate uh, you know, [9:19] narrative, but automate the whole thing, [9:21] saving the supervisor hours." So we took [9:25] a sample narrative that the supervisor [9:27] had written because we wanted the output [9:28] to be in kind of her style and wording [9:30] and format and we fed that into the [9:32] system to train it and then we iterated [9:34] through various formatting options with [9:36] the operations team and we considered [9:37] what time of day we would want to run [9:38] this thing so that the data was was [9:40] where we needed it to be and then we um [9:43] then we were able to actually go ahead [9:44] and and launch the thing. So now the [9:48] system automatically produces the [9:49] narrative and sends it to the supervisor [9:52] for a quick review and and some editing [9:54] if she needs to and then she can release [9:56] it to the technicians so that they know [9:57] their priorities for the day. And so on [10:00] the right you can see theuler data [10:02] output in Excel format and this was very [10:04] clunky difficult for the supervisors to [10:06] interpret and understand. Really theuler [10:08] on my team was the only one that could [10:10] read it and make you know sense of it. [10:12] And then on the left you can see the [10:13] daily schedule narrative which to all of [10:15] you probably looks mostly like gibberish [10:17] but but to the team in the actual work [10:19] site um it's formatted in a way that [10:21] they were very much accustomed to and [10:23] they can quickly read through it [10:24] understand their priorities and you know [10:26] what they needed to focus on uh for that [10:28] day but it's saving our supervisor hours [10:31] uh every day which is a huge [10:33] breakthrough. So um results um through [10:36] this effort we realized both hard and [10:39] soft results and some of them you know [10:41] as you can tell by the story we we [10:42] didn't really foresee when we began the [10:44] project we started with certain goals [10:46] and as we kind of learned more about the [10:47] platform and the capabilities some of [10:49] the goals changed and then we we figured [10:51] out how to do some much more compelling [10:53] things that we hadn't thought of. So [10:55] keeping our mind open to the new [10:56] technology really helped us break [10:57] through those kind of barriers of of you [10:59] know lack of understanding. Um, we [11:02] improve the accuracy of our schedule [11:04] while also making the output far more [11:06] accessible to the end users. You can [11:07] imagine if we were to send those Excel [11:09] files off to typical technician or [11:11] assembler, it's not going to mean much [11:13] to them. Uh, but now it gets to them in [11:15] an English language format that everyone [11:17] can interpret. Uh, we've also created [11:19] some compelling analytics to help us [11:21] improve our cycle times. In fact, right [11:22] now the team is actually using output [11:24] from Rapid Canvas to understand which [11:27] workstations we need to train more of [11:28] the technicians on. And this is going to [11:30] help us reduce overtime because we have [11:32] certain technicians that are very good [11:34] at what they do and in some cases may be [11:36] the only person that can do what they [11:37] can do and we don't want them having to [11:39] work every weekend and putting in all [11:41] these extra hours just so that we can [11:42] you know make sure the production stays [11:44] on track. Uh we we're going to save [11:46] hundreds of hours a year uh even just in [11:48] the supervisor time but also uh in um [11:52] you know giving people a quick way to [11:53] have those questions answered through [11:54] the chatbot. Uh that was something that [11:56] was uh pretty exciting and I know that's [11:58] only going to improve as we continue to [12:00] work on it and um improve the data and [12:02] the way that uh you know things are are [12:05] uh being given back to the person using [12:07] the the bot. And then finally we we [12:10] learned a lot about just AI data [12:12] platforms and how they can help a [12:13] manufacturing operation and and we have [12:15] a long to go but but now we can build on [12:17] this foundation and drive even more [12:18] results. [12:21] So if you find yourself um you know [12:24] embarking on a on a similar mission um [12:26] so I have a few tips to help you out if [12:28] you're going to try to do something like [12:29] this. So so the first thing it might [12:31] almost be stating the obvious but [12:32] centralize your data and try to get it [12:34] up into the cloud. Uh this is one of the [12:36] few things I probably would have done [12:37] differently had we brought our AI [12:39] partner in sooner. Um as you saw we we [12:42] started out with a local SQL server just [12:44] because that's what we were used to and [12:45] it was easy to get it up and running. [12:47] Um, if I was to do this over again, I [12:49] would probably be be pushing the data [12:51] directly into Snowflake and and skip a [12:53] lot of that. Um, so, uh, the other thing [12:56] is that it this makes it very easy to [12:58] connect the AI data platform. So, [13:00] basically within minutes of connecting [13:01] Rapid Canvas to Snowflake, I have [13:04] multiple people in my team, you know, [13:05] sitting around a table in a workshop [13:07] generating Python code through chat [13:09] prompts. And that was kind of the [13:10] funnest thing to see because I have [13:12] multiple people on the team that had [13:13] never used Python before. [13:15] And so it's all about speed to [13:17] opportunity. [13:19] Bring in your AI data partner as as [13:21] early as you can. Um even it's just to [13:24] assess your situation to see if their [13:26] platform will help. There's a lot they [13:27] can coach you on before you even spend a [13:30] dime. Uh things like, you know, getting [13:32] your data into the cloud and that sort [13:33] of thing. Um as you build, continually [13:37] checking in with your your end users. [13:39] The previous session had a great little [13:41] uh section where they talked about [13:42] change management. And you know, people [13:44] are intimidated by this stuff right now. [13:46] I mean, we have technicians in our line [13:47] who English is not their first language. [13:50] They um struggle sometimes to follow our [13:52] work instructions properly. And you [13:54] start throwing a bunch of crazy [13:55] high-tech at them. And people might kind [13:57] of, you know, pull back. So, um it they [14:01] might be reluctant to tell you, hey, [14:02] this fancy thing that you've built for [14:04] them isn't working so well. Uh so, check [14:06] in frequently. The the more you ask, the [14:07] more truth you're going to get. [14:10] And then uh finally um be ready to [14:12] really learn and invest in AI [14:14] capabilities. I don't have a data [14:15] scientist on my team and uh but I told [14:18] people I say hey this is a learning [14:20] opportunity you know let's give this a [14:21] try. Uh thankfully Rabbit Canvas does [14:23] have a data scientist and they were able [14:25] to help us. Um but you know have people [14:27] really ready to to try out these tools [14:29] and give them time and patience and some [14:31] space and for from my perspective don't [14:33] think of AI as a quick fix but more as a [14:35] way to lift the floor of your [14:37] capabilities. you'll be be able to [14:38] accomplish new things and in some cases [14:40] a lot faster than if you don't, you [14:42] know, do these things. [14:45] And finally, uh if you'd like to learn [14:46] more about Rapid Canvas, you can scan [14:48] this QR code, but um you might want to [14:50] hurry. So, thank you. [14:53] >> Thank you. So, uh I have a question [14:55] about your test deck. So are you go are [14:57] you using uh geni or just a machine [15:00] learning model because uh we have a [15:03] issue that when we use geni to predict [15:06] the actual number then sometimes it's [15:08] not very accuracy. So what do you think [15:11] about this? [15:12] >> I'm not I'm not sure I understand the [15:14] question. You're you're referring to [15:16] something I don't know I don't think. [15:18] >> So in your uh software architecture did [15:20] you use machine learning model or just [15:23] geni? [15:24] >> Oh yeah. Um, God, I don't know if I can [15:27] even answer that question. I may need [15:28] somebody from Rapid Canvas to answer [15:30] that. So, um, you should stop by their [15:32] booth and ask them what we're doing. [15:34] Yeah, it's So, I'm pretty sure we're [15:36] using large language model to produce [15:37] the the the daily narrative. Um, and [15:40] then obviously the LLM, I think, to to [15:42] assist with the using chat prompts to [15:44] create the code and the analytics, you [15:46] know, that type of thing. So, that's [15:48] probably a poor answer, but that's the [15:49] best I could do. Oh, yeah. So, the [15:52] question is, how large of a company is [15:53] Ksite? So we're about a $5 billion [15:55] company. Um we we primarily build [15:57] electronic test and measurement um you [15:59] know equipment. That's not what the site [16:01] I support does, but that's that's our [16:03] main business. And we have sites all [16:05] over the world, you know, Malaysia, US, [16:07] Europe.