Dataflix Cuts Delivery Time 31% With In-House AI Factory Framework

Dataflix依托自研AI工厂框架将交付时间缩短31% 将零散的生成式AI工具标准化为结构化的多智能体运营框架，使企业交付成本降低25%，QA工作量减少42%。 Srinivas Jayanthi详细介绍了Dataflix如何通过标准化的加速工具（包括Jo Prompts、Jo QA以及Tokconomics多云成本可观测性仪表板）取代零散的生成式AI应用，从而实现内部IT交付运营转型。他提出了涵盖基础（Foundation）、集成与知识（Integration & Knowledge）、智能（Intelligence）、体验（Experience）、运营（Operations）以及治理（Governance）六大支柱的Jo Factory框架，以构建企业级AI交付体系。Jayanthi重点展示了交付周期缩短31%、运营成本降低25%等内部指标，并分享了一个汽车召回的实际案例——该案例实现了73%的保修自动验证，周转速度提升了16倍。 面对客户要求在两个月内低成本交付的严苛需求，Dataflix 从零散使用 Cursor 和 Claude 转向系统化的内部 AI 交付体系改造。 (0:04) Dataflix 针对核心工作流开发了四大内部加速器：管理通用提示词的 Jo Prompts、自动生成技术蓝图的 Jo 架构 Agent、自动化测试的 Jo QA 以及统一多云 Token 成本监控的 Tokconomics。 (3:04) 基于内部实践，Dataflix 提炼出涵盖基础、集成知识、智能层、体验层、运维及治理的六支柱 Jo Factory 框架，并将多云 AI 成本分析平台 Tokconomics 免费向公众开放。 (7:48) 系统化引入 AI 后，公司内部设计返工减少 38%，QA 工作量降低 42%，整体交付提速 31%，并成功将 Agent 应用扩展到招聘、销售以及负责初审保密协议的 Jo Legal。 (13:24) 在汽车主机厂召回理赔落地案例中，智能视觉 Agent 实现了 73% 的维修照片自动审核，将处理速度提升 16 倍，并减少了 81% 的经销商跟进沟通。 (16:14)

Srinivas Jayanthi, Chief Technology Officer at Dataflix, outlines how the IT services firm formalized its internal generative AI usage into dedicated accelerators to satisfy client demands for shorter delivery cycles and reduced costs. This internal initiative formed the basis of the Jo Factory Framework, a structured six-pillar architecture designed to deploy agentic systems across software engineering and enterprise operations. Engineering Standardization and Multi-Cloud Token Observability (3:04) To address client demands for two-month turnaround times and lower budgets, Dataflix replaced ad hoc engineer usage of tools like Cursor, Claude, and OpenAI with standardized internal accelerators across four key software delivery domains. The company introduced Jo Prompts to centralize and reuse prompt patterns for standard features like single sign-on implementations, eliminating redundant prompt drafting. To standardize solution design, Dataflix introduced an architecture agent that ingests business requirements and generates standardized solution architecture blueprints with guardrails to constrain technology stack choices. In testing, the firm deployed Jo QA to parse business requirements documents and automatically produce structured test plans, test cases, and automated execution scripts. To address token usage across 10 to 12 distinct model and cloud vendors—including AWS, Azure, GCP, Cursor, and Anthropic—Dataflix developed Tokconomics. The tool normalizes token telemetry into a single data model, tracking multi-cloud AI expenditures and cost per million tokens across individual users, projects, and departments. The Six-Pillar Jo Factory Architecture (7:48) Dataflix organized its internal accelerators into three operational buckets—Engineering, Operations, and Experience—which it formalized as the Jo Factory Framework. The framework consists of four sequential delivery stages: Foundation (evaluating infrastructure, APIs, and data readiness), Integration and Knowledge (building semantic context layers), Intelligence (orchestrating agent workflows), and Experience (interfacing through web, chat, voice, or IDE environments). Operating across all four stages are automated CI/CD pipelines under Operations and human-in-the-loop compliance and model explainability under Governance. Rather than distributing these tools as a proprietary product suite, Dataflix uses the framework as delivery accelerators for client engagements. As part of this initiative, the firm released its Tokconomics cost observability dashboard for free public use, allowing external organizations to monitor multi-provider token consumption and cost unit metrics by integrating their API keys. Cross-Functional Agent Adoption and Internal Performance Metrics (13:24) Dataflix extended agentic automation beyond software engineering into general business operations. The firm deployed agents in human resources for candidate screening and interview workflows, and in sales for digital outreach and introductory customer interactions. For routine legal review, Dataflix developed Jo Legal, an agent trained on company agreement templates to review initial non-disclosure agreements and preliminary customer redlines, reserving external legal counsel for complex Master Services Agreements and liability terms. According to Jayanthi, the internal implementation of these accelerators produced quantifiable delivery gains: a 38% decrease in architectural design rework, a 42% reduction in QA effort, 35% fewer change requests and production escape defects, 31% faster project completion timelines, and an overall 25% reduction in operational running costs. Production Deployment in Automotive Warranty Inspections (16:14) To demonstrate the Jo Factory architecture in production, Jayanthi highlighted an implementation for an automotive Original Equipment Manufacturer (OEM) managing vehicle safety recalls across dealership networks. Dealership technicians historically repaired components, captured physical photographs of the replacement, and submitted documentation for manual OEM warranty claim review and reimbursement. Dataflix implemented vision and document agents to automate inspection by verifying repair photos against standard operating procedures, validating replacement parts, and checking for complete metadata. The automated warranty inspection pipeline yielded a 73% automated image validation rate without human intervention, an 81% reduction in dealership follow-up requests regarding incomplete documentation, and a 16-fold increase in turnaround speed compared to manual processing. Jayanthi noted similar multi-agent systems are operating across fraud detection, IT support desks, HR assistance, hospitality services, and supply chain inventory management. Dataflix argues that enterprise AI adoption requires shifting from ad hoc developer experimentation to structured, observable frameworks with strict architectural and cost guardrails. However, the firm notes that production multi-agent workflows still rely fundamentally on human-in-the-loop oversight to resolve edge-case discrepancies, legal liability negotiations, and complex exceptions.

Dataflix 首席技术官 Srinivas Jayanthi（JV）指出，面对客户对项目交付周期缩短至两个月以内及削减预算的严苛要求，公司通过内部推行“Jo”系列 AI 加速器与六支柱架构，实现了从零散工具使用向标准化、企业级多智能体工作流的系统化转型。 从零散生成工具到四大标准化研发加速器 (3:04) 在传统敏捷开发模式下，Dataflix 工程师曾零星使用 Claude、Cursor 及 OpenAI 编写代码，但工具使用高度分散且缺乏规范。为了在时间和成本压力下提高交付规模，公司梳理出四大核心实践领域并打造了针对性加速器：通过 Jo Prompts 集中管理与复用高频提示词（如 Google 与 Microsoft 单点登录），消除重复劳动；利用 Jo Architecture Agent 自动将业务需求转化为带有安全边界的标准化系统架构蓝图，避免技术栈混乱；借助 Jo QA 直接解析业务需求文档（BRD），自动生成测试用例与执行脚本；同时搭建多云成本监控平台 Tokconomics，统一接入 AWS、Azure、GCP、Cursor 及 Anthropic 等 10 至 12 家供应商的 Token 账单，提供按部门、项目与人员划分的单价与消耗视图。 Jo Factory 六支柱企业级 AI 架构框架 (7:48) 通过将内部加速器归纳为工程、运营与体验三大维度，Dataflix 提炼出 Jo Factory 框架，为规模化构建企业级 AI 系统提供标准蓝图。该框架包含四大核心交付层：首先是评估数据就绪度、API 可用性及底层云基础的“基础层”（Foundation）；其次是构建语义环境与企业上下文的“集成与知识层”（Integration & Knowledge）；第三层是用于智能体编排的“智能层”（Intelligence Layer）；最终通过网页、对话、语音或 IDE 工具交付给用户的“体验层”（Experience Layer）。贯穿交付全流程的则是负责自动化流水线的“运营层”（Operations）以及确保模型可解释性与人工在环验证的“治理层”（Governance）。Dataflix 将这套工具作为实施加速组件，并将 Tokconomics 平台开放供公众免费接入使用。 跨业务运营拓展与内部效能提升成效 (13:24) 在软件工程之外，智能体自动化进一步延伸至人力资源、销售与法务等业务环节：HR 智能体负责初筛与面试，数字化销售智能体开展电话外呼与业务推介，Jo Legal 则基于企业历史合同与条款库自动审核初步保密协议（NDA）及修改标记，仅将复杂的专业主服务协议（MSA）移交外部法律顾问。根据 Dataflix 披露的内部运营数据，这一系列系统化举措使系统设计返工率降低了 38%，质量保证（QA）工作量减少了 42%，变更请求与生产环境遗留缺陷减少了 35%，项目整体交付周期提速 31%，并将运营运行成本削减了 25%，为组织在预算紧缩下扩充产能提供了实际验证。 汽车召回检验落地案例与多智能体生态 (16:14) 在外部企业落地方案中，某汽车原始设备制造商（OEM）借助视觉与文档智能体改造了涉及数千家经销商和数百万辆召回车辆的保修审核流程。传统流程依赖技师维修后拍照并上传人工审核，而部署的智能体能够根据标准操作规程自动比对零件安装照片并验证元数据。实际运行数据显示，该方案实现了 73% 的图像自动化验证且无需人工干预，索赔处理周转速度提升至传统人工流程的 16 倍，并在提交环节即时纠错，将经销商后续跟进沟通减少了 81%，异常情况则交由人工复核。此外，该技术架构已拓展至防欺诈审查、IT 服务台协助以及供应链多仓库存调配等多智能体并行场景。 Dataflix 的实践表明，服务机构应对交付周期的缩减不能仅停留在孤立的 AI 编码辅助，而需建立覆盖数据层、架构设计、Token 成本核算与人工在环治理的完整交付体系；不过，复杂主服务协议谈判等高风险环节仍需依赖人工专业判断。

## Background 该内容由 Ai4 频道在 YouTube 平台上发布，主讲人为 Dataflix 首席技术官 Srinivas Jayanthi（JV）。视频简介表明，演讲围绕企业级 AI 平台 Dataflix Jo 及其实践蓝图“AI 工厂框架”（AI Factory Framework）展开，详细阐述了企业如何通过连接数据、模型与体验层的统一架构，系统化地跨越从实验原型到规模化生产级 AI 的落地鸿沟，并实现成本控制与安全治理。 ## TL;DR Dataflix CTO Srinivas Jayanthi 分享了公司面对客户交付周期紧迫与预算削减的双重压力下，通过自身 AI 转型孵化出企业级 AI 平台 Jo 的历程。团队在内部梳理并沉淀出 Prompt 资产库、架构设计智能体、Jo QA 自动化测试以及 Tokconomics 跨云成本观测工具四大加速器，并提出了涵盖基础设施、知识、智能编排、体验、运营与治理的六支柱“AI 工厂”框架。此外，演讲还展示了智能体在法务、销售、HR 等非技术部门的落地成效（如交付速度提升 31%、QA 工时减少 42%），并通过汽车缺陷召回自动化审核（效率提升 16 倍）等实际案例展示了多场景智能体协同运营的巨大价值。 ## Atmosphere & style 整场演讲风格务实干练、极具工程落地质感，讲者以亲和幽默的“JV”自居，开场便以行业交付痛点与法务环节的现实调侃拉近与观众的距离。叙事逻辑清晰连贯，从工程师日常工具链的混乱切肤之痛层层递进到严密的平台架构设计与详实的 ROI 数据验证。现场氛围既展现了资深技术架构师在成本控制与架构治理上的严谨克制，又传递出对 AI 智能体深度嵌入企业微循环、重塑产业生产力的坚定信心与充沛活力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=4s) Dataflix 孵化企业级 AI 平台 Jo 的初衷是应对交付周期紧迫与预算大幅缩减的双重压力，先在组织内部实现 AI 驱动的生产范式转型。 - [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) 针对工程师使用 AI 工具零散且缺乏系统化治理的痛点，Dataflix 在内部提炼并构建了涵盖 Prompt 治理、架构设计、测试自动化与成本监控的四大加速器。 - [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) Jo Prompts 建立了提示词与代码资产管理体系，实现常用功能生成逻辑的高质量沉淀与全局复用。 - [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) Jo QA 智能体能够直接解析业务需求文档（BRD），端到端自动化生成标准测试用例、测试计划及执行脚本。 - [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) 针对接入 10-12 个不同模型供应商带来的成本失控风险，Dataflix 打造了 Tokconomics 平台，通过统一数据模型实现跨云跨模型的 Token 成本可观测性。 - [07:26](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=446s) Jo Factory 框架提出了包含底座、知识、智能编排、交互体验，以及运营运维与治理支撑的六支柱系统化工程蓝图。 - [07:26](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=446s) Tokconomics 成本分析工具对外免费开放，帮助企业清晰看清不同供应商的 AI 真实花销，消除规模化采用的阻碍。 - [13:24](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=804s) AI 智能体已全面渗透至非技术职能，例如 Jo Legal 独立承担前期 NDA 和法务审查，数字销售与 HR 智能体接管商机初筛与第一轮面试。 - [13:24](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=804s) 内部量化数据显示，AI 实践使设计返工率降低 38%、测试工时减少 42%、项目交付速度提升 31%，同时系统运行成本降低 25%。 - [16:14](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=974s) 在汽车制造业召回理赔场景中，通过图像审核智能体与 Human-in-the-Loop 人机协同机制，实现了 73% 的自动验证覆盖率并使业务周转效率提升 16 倍。 - [16:14](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=974s) Jo 平台已支撑金融安全防欺诈、IT Helpdesk 故障排查、酒店虚拟礼宾及多仓供应链调度等多样化工业与企业级智能体落地。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=4s) 缘起与背景：Dataflix 的自身 AI 转型之路 Dataflix 首席技术官 Srinivas Jayanthi（大家都称呼他为“JV”，因为家乡重名的人太多而保留了这个简短的昵称）在开场时介绍了公司的定位与使命。Dataflix 总部位于旧金山湾区，并在美国、墨西哥和印度设有分支机构，是一家专注于端到端人工智能解决方案的公司。团队的工作涵盖了从底层的底层数据就绪（Data Layer）一直延伸到最终用户的交互体验层（Experience Layer）。在为各类企业客户构建大量 AI 智能体（AI Agents）与现代化应用的过程中，团队产生了一个根本性的思考：作为一家输出 AI 能力的技术公司，Dataflix 自身究竟该如何利用 AI 实现内部生产力的重构与转型？ 正是为了回答“如何先武装自己、再赋能客户”这一命题，Dataflix 最终孵化出了企业级 AI 平台——Jo 平台。JV 指出，传统的 IT 交付模式依靠成熟的敏捷开发方法论（两周一个 Sprint 迭代、梳理业务需求文档等），虽然能够按部就班地交付优质系统，但如今客户的需求和市场节奏发生了剧烈变化。“更快”与“更便宜”成为了所有客户的核心诉求——无论是希望交付周期压缩到两个月内，还是面临大幅削减的预算压力。大家都在追问同一个问题：“我们能否用更少的成本做更多的事（Do more for less）？”面对这一挑战，Dataflix 决定从自身组织内部动刀，系统性探索 AI 驱动的生产范式。 - **公司布局**：总部设于旧金山湾区，交付网络覆盖美国、墨西哥及印度。 - **业务范围**：数据治理、应用现代化改造、全栈 AI 解决方案及智能体构建。 - **转型驱动力**：应对交付周期紧迫（Time Pressure）与预算缩减（Cost Pressure）的双重压力。 --- ### [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) 告别零散实验：提炼内部四大 AI 加速实践 在内部审视中，团队发现工程师们其实早已在广泛使用 AI——大家日常在 Claude、Cursor、OpenAI 等各种工具中编写代码、调试逻辑。然而，这种使用方式是零散、孤立且缺乏系统化治理的（"there was no systematic and structured way of handling it internally"）。为了让 AI 真正转化为组织级工程能力，Dataflix 设定了明确的目标，并在内部最密集使用 AI 的四个关键实践领域开发了针对性的“加速器”（Accelerators）： 1. **Prompt 治理与复用（Jo Prompts）**：团队发现不同项目的工程师经常为了相同的功能（例如基于微软或谷歌账号的第三方登录 SSO）重复编写和调试未经优化的 Prompt，甚至反复生成相似的代码。为此，Dataflix 建立了提示词资产管理平台，实现高质量 Prompt 与生成逻辑的沉淀和全局复用。 2. **架构与方案设计标准化**：面对不同项目技术栈碎片化的问题，团队训练了具备工程护栏（Guardrails）的设计智能体。工程师只需输入业务需求，经过交互式澄清，智能体即可自动输出标准化的系统架构、技术选型建议与解决方案设计文档。 3. **QA 测试自动化（Jo QA）**：针对测试这一工程中的“低垂果实”，构建了 Jo QA 智能体。该工具能直接解析业务需求文档（BRD），端到端生成详尽的测试用例、测试计划及自动化执行脚本，不仅提升了效率，更确保了全团队脚本编写规范的一致性。 4. **AI 成本可观测性（Tokconomics）**：随着 AWS、Azure、GCP 以及 Anthropic、Cursor 等 10 到 12 个不同 API/模型供应商的接入，各供应商 Token 计量方式五花八门，AI 算力成本面临失控风险（"AI costs, and if you don't pay attention to it, it can spiral out of control"）。Dataflix 构建了统一数据模型（Common Data Model），将跨云、跨模型的消耗数据汇聚成多维度看板，精确核算每百万 Token、每个项目、每位用户乃至每个部门的单价成本。 ```mermaid flowchart TD accTitle: Dataflix Internal AI Accelerators subgraph S1["输入与设计阶段"] REQ["业务需求 / BRD"] --> DES["方案设计智能体<br/>(标准架构/技术栈)"] PRM["Jo Prompts<br/>(通用 Prompt/代码库)"] --> DES end subgraph S2["研发与测试阶段"] DES --> DEV["工程编码交付"] REQ --> JQA["Jo QA 自动化智能体<br/>(测试用例与执行脚本)"] DEV --> JQA end subgraph S3["运营与成本治理"] DEV --> TOK["Tokconomics 成本可观测平台<br/>(汇聚 AWS/Azure/GCP/Anthropic 数据)"] JQA --> TOK end ``` --- ### [07:26](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=446s) Jo Factory 架构蓝图与 Tokconomics 开放体系 随着内部实践的深入，Dataflix 总结出所有 AI 加速器其实都归拢在三个核心维度：**工程层（Engineering）**（解决如何构建）、**运营层（Operations）**（解决监控与运维）和**体验层（Experience）**。JV 特别强调了体验层的重要性，因为不同的角色感知 AI 的形态完全不同：工程师面对的是 IDE 插件，业务人员面对的是分析看板，最终用户面对的则是聊天助手或语音界面。 为了将这些加速模块统一起来，Dataflix 提出了六支柱的 **Jo Factory 框架**（Jo Factory Framework），作为企业 AI 落地的系统化工程蓝图： ```mermaid flowchart LR accTitle: Jo Factory Framework 6-Pillar Blueprint subgraph Foundation["1. 基础设施底座 (Foundation)"] F1["数据就绪 / API / 基础算力设施"] end subgraph Knowledge["2. 知识与上下文 (Knowledge)"] K1["多源数据集成 / 语义与上下文层"] end subgraph Intelligence["3. 智能编排 (Intelligence)"] I1["多智能体编排 / 流程调度"] end subgraph Experience["4. 交互体验 (Experience)"] E1["Web 应用 / 聊天机器人 / 语音 / 看板 / IDE"] end Foundation --> Knowledge --> Intelligence --> Experience subgraph GovernanceOps["横向支撑体系"] OPS["5. 运营运维 (Operations)<br/>CI/CD 流水线 / 监控调度"] GOV["6. 治理与解释性 (Governance)<br/>可解释性 / Human-in-the-Loop"] end GovernanceOps -.-> Foundation GovernanceOps -.-> Knowledge GovernanceOps -.-> Intelligence GovernanceOps -.-> Experience ``` 在现场演示中，JV 重点展示了已经对外免费开放的成本分析工具 Tokconomics。用户只需接入自己的 API 凭据或云服务源，系统即可自动对齐数据并生成多维分析报表。JV 直言团队并不打算将该工具单独打包成高价软件出售，而是作为开源/免费加速组件直接交付给客户：“我们的‘自私目的’很简单，就是让每个人都能清晰看清跨不同供应商的 AI 真实花销，从而加速实际业务价值的达成。” --- ### [13:24](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=804s) 跨业务部门落地与量化降本增效成果 AI 的应用远不止于纯技术研发团队，Dataflix 已经在人力资源、财务、销售和法务等多个非技术职能中全面部署了定制智能体。JV 幽默地提醒现场的律师朋友“请先捂上耳朵”——在日常商务往来中，绝大多数初期的客户互动都会涉及保密协议（NDA）或基础文书审查，以往事事依赖外部法务不仅耗时长且费用昂贵。为此，团队打造了名为 **Jo Legal** 的智能体，其在充分学习公司法律条款与合规偏好后，能够独立承担前期的法务审查工作；只有在推进到主服务协议（MSA）等涉及复杂合同责任的阶段时，才会引入专业律师团队。 在业务拓展方面，Dataflix 部署了数字销售智能体（Digital Sales Agents），能够自主拨打潜在客户电话、清晰介绍公司业务并进行初步商机筛选；在 HR 场景中，智能体则接管了候选人的初筛和第一轮结构化面试。这些实践证明，AI 智能体已经深入到业务运营的各个微循环中。 针对这些举措带来的实际成效，JV 强调“空口无凭，事实胜于雄辩（Proof of the pudding is in its taste）”，并公布了内部沉淀的核心量化数据： - **设计返工率**：减少 **38%** - **QA 测试工时**：降低 **42%** - **变更请求（CR）及逃逸缺陷**：减少约 **35%** - **整体项目交付速度**：提升 **31%** - **系统运行时成本（Run Cost）**：降低 **25%** --- ### [16:14](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=974s) 汽车行业召回案例与全场景智能体实践 为了展现 Jo 平台在外部复杂产业环境中的承载能力，JV 分享了一个汽车制造业的实际客户案例：该项目面临数千家汽车经销商（Dealerships）以及涉及数百万辆汽车的缺陷召回（Recall Operations）处理。 在传统的召回流程中，经销商对召回车辆完成维修后，必须拍照并上传工单及图片证明；随后汽车主机厂（OEM）需要安排大量人工核对照片、审查保修索赔单据。Dataflix 联合客户构建了自动化图像审核与理赔处理智能体：智能体自动对经销商上传的维修照片进行特征与质量校验，并核对单据一致性；只有在检测到异常或存在存疑分歧时，才会自动触发“人机协同”（Human-in-the-Loop）机制转交人工介入。 ```mermaid flowchart TD accTitle: Automotive Recall AI Workflow D["经销商完成维修并上传现场照片"] --> AGT["AI 验证智能体<br/>(图像特征识别 + 单据校验)"] AGT -- 验证通过 --> OEM["自动核准索赔并进入结算"] AGT -- 发现异常 / 缺失材料 --> HIL{"人工介入审批<br/>(Human-in-the-Loop)"} HIL -- 驳回并提示补件 --> D HIL -- 人工确认无误 --> OEM ``` 该系统上线后取得了显著的业务改善： - **图像自动验证覆盖率**：达到 **73%** - **业务周转效率（Turnaround Time）**：相比纯人工流程提升了 **16 倍** - **经销商补件与重复跟进（Follow-ups）**：减少了 **81%** 在演讲的尾声，JV 展示了一组实时运行的智能体业务监控看板录像，呈现了多种多样的工业与企业级场景：包括在后台实时拦截欺诈行为的金融安全智能体、自动分类并解决 IT 故障的 IT Helpdesk 智能体、HR 咨询助手、酒店与文旅行业的虚拟礼宾接待（Concierge & Guest Services），以及能够跨多个区域仓库智能决策零部件调拨路径的供应链调度智能体。 JV 总结道，企业利用 AI 打造竞争壁垒是一场持续进化的旅程，Dataflix 已经通过 Jo 平台在内部完成了全流程的方法论验证，并期待将这套成熟的“AI 工厂”能力赋能给更多寻求规模化 AI 落地的企业伙伴。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=4s) faster and cheaper is the theme that we hear from everybody — Srinivas Jayanthi > [00:04](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=4s) we started this journey internally to see what can we do to transform the way we work, the way we build systems — Srinivas Jayanthi > [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) there was no systematic and structured way of handling it internally — Srinivas Jayanthi > [02:50](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=170s) AI costs, and if you don't pay attention to it, it can spiral out of control — Srinivas Jayanthi > [07:26](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=446s) Tokconomics gives you a single pane of glass view across all these different LLMs, different models, different providers — Srinivas Jayanthi > [07:26](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=446s) our selfish goal is to make it easy for everyone to see what is happening across different vendors — Srinivas Jayanthi > [13:24](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=804s) lawyers in the room, please close your ears for a second — Srinivas Jayanthi > [13:24](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=804s) the proof of the pudding is in its taste — Srinivas Jayanthi > [16:14](https://www.youtube.com/watch?v=BnTNpP3cpSc&t=974s) this is a journey, and every company has to go through this journey to stay competitive and stay relevant — Srinivas Jayanthi ## Mentioned Srinivas Jayanthi (JV), Dataflix, Dataflix Jo, AI Factory Framework, Jo Prompts, Jo QA, Tokconomics, Jo Legal, Claude, Cursor, OpenAI, AWS, Azure, GCP, Anthropic ## Video description Srinivas Jayanthi - Chief Technology Officer, Dataflix Dataflix Jo is an enterprise AI platform built to accelerate AI execution at scale, transforming how organizations move from idea to production. At the heart of Jo is the AI Factory Framework, a practical blueprint designed to help organizations move from experimentation to scalable, production-grade AI. This session will reveal how a structured “AI Factory” approach enables teams to systematically plan, build, scale, and govern AI solutions, connecting data, models, and user experiences across a unified architecture. From foundational data readiness to agentic orchestration and real-world copilots, discover how to transform fragmented capabilities into a cohesive, high-impact AI delivery engine, while ensuring cost control, security, and continuous improvement at every stage. [Source](https://www.youtube.com/watch?v=BnTNpP3cpSc)

[0:04] Srinivas Jayanthi: faster and cheaper is the theme that we hear from everybody [0:04] Srinivas Jayanthi: we started this journey internally to see what can we do to transform the way we work, the way we build systems [2:50] Srinivas Jayanthi: there was no systematic and structured way of handling it internally [2:50] Srinivas Jayanthi: AI costs, and if you don't pay attention to it, it can spiral out of control [7:26] Srinivas Jayanthi: Tokconomics gives you a single pane of glass view across all these different LLMs, different models, different providers [7:26] Srinivas Jayanthi: our selfish goal is to make it easy for everyone to see what is happening across different vendors [13:24] Srinivas Jayanthi: lawyers in the room, please close your ears for a second [13:24] Srinivas Jayanthi: the proof of the pudding is in its taste [16:14] Srinivas Jayanthi: this is a journey, and every company has to go through this journey to stay competitive and stay relevant

[0:04] Srinivas Jayanthi: JV [0:04] Srinivas Jayanthi: leverage AI internally [0:04] Srinivas Jayanthi: Can we do more for less? [3:04] Srinivas Jayanthi: Jo Prompts [3:04] Srinivas Jayanthi: Jo QA [3:04] Srinivas Jayanthi: Tokconomics [7:48] Srinivas Jayanthi: Engineering [7:48] Srinivas Jayanthi: Operations [7:48] Srinivas Jayanthi: Experience [7:48] Srinivas Jayanthi: Jo Factory Framework [13:24] Srinivas Jayanthi: please close your ears [13:24] Srinivas Jayanthi: Jo Legal [13:24] Srinivas Jayanthi: do more for less

[0:04] Well, thank you. My name is Shwinas Jayanti.  Uh, you can call me JV. Those are my initials. [0:09] The place where I come from, quite a few shinies  around. So, somehow my initials stuck with me. Um, [0:16] so I'm the chief technology officer at Dataflex.  We are an AI solutions firm. uh we build a lot of [0:25] AI solutions to different customers based on their  needs. We have we go right from the data layer all [0:32] the way up until the experience layer how people  experience AI and and whatnot. And through this [0:38] journey we build a lot of AI agents. And uh  the story here is what did we do in order to [0:46] leverage AI internally, right? In terms of uh how  did we transform ourselves and this is really our [0:53] story and uh you you know hopefully you pick a  few things out out of this. Uh so who are we? [1:01] Well we are an AI solutions firm. We are located  in we are headquartered in the Bay Area. We uh are [1:07] in the US, Mexico and in India. And in terms of uh  what we do build a lot of as I said AI solutions, [1:15] we work with data, we modernize apps and through  this journey we ended up building a platform to [1:21] help ourself help you and that's the Joe platform.  So the story here is about how we came up with [1:27] this Joe platform, how has it helped us and  how can we leverage that to help you right? So [1:35] where did we really start right? We let me  take you back to where we really uh started [1:41] this entire journey. Um we as I said build a lot  of solutions. We are an IT services firm and we do [1:49] this really well. We uh have a tried and tested  mechanism. We all know agile two week sprints [1:57] well definfined requirements sometimes not  so well defined. How do we handle that? [2:01] Yeah, we have done that and we have delivered  some good solutions along the way. But what [2:06] we realized was the growing challenge with a  lot of our customers in terms of their demand, [2:12] right? What is it that they are looking for and  cost pressure was one hey I need this faster [2:18] and I need it now or I need it in 2 months time.  So essentially one was time pressure the other [2:25] was cost. They wanted it at a much lower cost. So  the question that we all asked ourselves was well [2:32] essentially what you're looking for is can we do  more for less? we we're going to spend the same [2:38] amount of money maybe less but can you do more  for the same amount right and and that's where [2:44] we started thinking and internalizing it  the question that we asked ourselves right [2:50] and that's where the AI factor started coming  in and um so what we realized uh we are doing [2:57] so much of this work for our customers but how  much are we actually doing this for ourselves [3:04] and when we started looking internally We started  uh seeing that yeah there were AI was being used [3:10] widely. We used that for writing code all of  you are probably familiar with. We used that [3:17] sporadically in a few different places uh claude  cursor openai you you name it you know we probably [3:23] have tried but there was no systematic and  structured way of handling it internally. [3:29] And so what we ended up doing was setting  ourselves some goals in terms of what we could do. [3:35] Uh so there were two organization goals that we  uh really started started putting together. And [3:44] in terms of uh how we went about uh accelerating  this, if you if you really look at it, [3:52] there were four different areas where we  ourselves were starting to use AI prolifically, [3:59] right? One was prompt management. We said, look,  this is an area where everyone is using a chatbot, [4:05] everyone is using cloud or some form in  which they are interacting with AI agents. [4:10] uh and the mode of interaction was always prompts.  So is there something that we can do there? [4:17] And so that's where prompt management became  paramount for us and there were different projects [4:21] using the same sort of prompts and they were all  less optimized. So we said how can we reuse and [4:27] and all that good stuff. So we ended up creating  uh a sort of accelerator for ourselves saying [4:35] if project A goes ahead and uses a certain prompt  to build an application. Let's take an example of [4:41] I want to log in using Microsoft or Google  or whatever, right? It's a common feature [4:50] that is there for most projects. If we are  using a prompt to go ahead and do that, [4:56] why are we redoing the same prompt again and  again? And it goes even further. What does the [5:02] prompt generate? Well, if it's generating code,  then why am I actually going ahead and generating [5:06] the same code again and again? why can't I  reuse that? So essentially it started getting [5:11] into that sort of conversation where we ended up  creating Joe prompts for prompt management. Then [5:19] we started looking at how are applications  being designed in the first place. Um and every [5:26] customer requires a different set of designs. uh  and internally when we started looking at it we [5:33] realized for all the internal projects  we were different projects were using [5:37] different stacks and whatnot. So we created  guardrails and we created an agent that can [5:44] help standardize how we go about creating these  solution designs. So you feed the requirement [5:51] at the end of an interactive process. The agent  spits out well-defined solution design in terms [5:57] of where you should go, what kind of stack you  should use and and and all that. So the second one [6:03] which we looked at was QA practice and uh from a  QA practice, it's a lowhanging fruit. We all know [6:09] that you know you could go about generating QA  test cases, automating it. So we ended up creating [6:15] Joe QA you know essentially trying to figure out  how you can automate testing based on business [6:20] requirements and and and whatnot. And lastly we  started uh also looking at cost observability. [6:27] You know all this is great in terms of what we  are doing for addressing specific practice areas [6:34] but the challenge was always with cost  and as most of you probably realize by now [6:40] you know AI call AI costs and if you don't pay  attention to it it can spiral out of control. [6:47] So what we really were trying to do out here was  bring about a common pain. So we use AWS, we use [6:53] Azure, we use GCP, we use cursor, Anthropic, you  name it. You probably have 10 or 12 sources out [7:00] there. Everyone has their own way of uh you know  accounting for the tokens and whatnot. So what we [7:05] did was we ended up building a common data model  to bring all of that data and then essentially [7:11] building our dashboards on top that gave birth to  some cost observability which we ended up calling [7:17] tokconomics. So what this prompted us to do  really was uh u to start thinking about it [7:26] in a much more structured manner. As much as we  did this internally, we realized these were we [7:33] started observing a pattern. The pattern was uh  all these accelerators that we are beginning to [7:40] address though we called them practice areas  they all followed three distinct buckets. [7:48] One was primarily around engineering. You know  what are we doing at an engineering level? How are [7:54] we actually building it? The second one was around  operations. And the third one was experience. [8:01] Experience was extremely important because  depending upon who you are as a user, [8:06] you would experience AI in your own way. If  you're an engineer, you are probably looking [8:11] at an ID. If you are an end user, you are probably  looking at a dashboard or some sort of solution. [8:17] So your experience could vary based on who you  are. So we started classifying all that we were [8:23] doing against these three buckets engineering  operations for monitoring and then experience [8:30] which took us to uh a thought around okay if  we are able to structure it why can't we go [8:38] about building axel radars that are reusable and  already the work that we did on our practice areas [8:46] helped us so whatever we did with the QA practice  ended up being Joe QA which was an axel data [8:53] that every one of in my team uses as they go  about implementing solutions what does it do [8:58] you feed a BRD again at a high level 50,000 ft  bird's eye view there there are several moving [9:04] components underneath you pass in a BRD out comes  a wonderful QA document uh including automated [9:12] testing scripts and all that good stuff right  so essentially you are trying to automate the [9:17] the key here is not just automation It's  consistent automation. It's consistency in [9:22] terms of how we go about writing scripts and and  whatnot. The another of these axelars if you see [9:29] is tokconomics that I was talking about. We ended  up creating uh an entire dashboard uh a system [9:37] that is now open to everybody. Uh if you want you  can try it yourself. Uh but essentially what it [9:44] does is import your sources, start visualizing  all the data and give you per unit metrics. [9:51] What is the cost of a million tokens? What is  the cost per project? What is the cost per user? [9:57] What is the cost per department? You know, these  are common. And then when you multiply that with [10:02] different sources, it becomes a bit more hairy  and complicated. So what we are trying to do [10:08] is just solve that. We are not selling this as  a product. We are we are essentially offering [10:13] these as accelerators that can fasten the time to  value for you right and another one is Joeactory [10:20] that I was talking about from an architecture  and design standpoint. So these are again not [10:24] to go through all nine of them but these are the  accelators that we have come up with so far that [10:30] we are using them internally and now we have begun  to use them for all the external projects also [10:37] but every one of them is rooted in what we call  the Joe factory framework that's the engineering [10:45] blueprint based on which we operate. So what  you see here is is is a six-pillar uh framework [10:54] where we start looking at what is your  foundation. Are you ready for an AI project or [11:00] do you have your data ready? Do you have your APIs  ready? Is your infrastructure ready and so on. [11:05] Then on top of that we start looking at  integration with all these data sources [11:09] and building the knowledge layer on top. Right?  This is where your semantics start uh start stuff [11:14] starts coming in. How do you build the context?  And then on top of that is the actual intelligence [11:20] layer where you start working with different types  of agents to help orchestrate your flows, plugging [11:27] in the context, plugging in the intelligence  and whatnot. And lastly is your experience. [11:32] How do you now experience AI? Right? You could  probably build a web app. It could be a chatbot. [11:38] It chat assistant or a chat agent or UI. It could  be a voice agent or it could be a a dashboard, [11:47] it could be an ID, whatever may be the case  depending upon who the user is. So we end up you [11:52] know with that and lastly all these go hand inhand  parallelly with operations where you talk about CD [11:59] pipelines and whatnot and then on the other hand  we talk about uh governance uh is your AI whatever [12:06] you're building capable of explaining itself.  So a lot of human in the loop comes in there. [12:14] So an example of what you're seeing here is uh let  me just run it. Uh this is tokconomics for you. [12:23] So when I talk about the accelerators that  we have built that you can access yourself. [12:29] This is an example. So you can simply plug in  your sources and once your source is plugged in, [12:35] you start visualizing all your data. It's as  simple as that. And I welcome you to try it out. [12:42] At the end of this presentation, there's a QR code  you can you can scan and you can simply log in and [12:47] create an account for yourself. And then this  is at no cost, right? You can try it yourself. [12:52] Our selfish intent here is every one of us has  to have visibility into what is our AI spend [13:01] across all the sources, right? And and uh this  really helps. So when we go about building [13:07] solutions for our customers, this is one of the  accelerators that we give to customers at no cost [13:15] so that they can use it themselves. Right now  all this is good. Uh so where are we headed? [13:24] Beyond just the engineering aspect of it, we  have started using it in several areas. Um [13:31] and and if you see we uh for instance people in  HR uh today we use agents to interview people [13:42] do all the screening the initial screening we  have started using that uh in terms of uh actual [13:50] operations we have started using that in  finance we have started uh leveraging that [13:55] a simple example I'll tell you legal I don't  know how many of you are attorneys So please [14:00] close your ears when I say this. But the point  is at least the first few interactions with uh [14:06] a customer always is through an NDA or basic  paperwork and all that for everything I can't [14:12] run back to my attorney. Uh it's going to take  time. It's going to be expensive and whatnot. So [14:17] what we do is we have you know an accelerator  called Joe legal where which is trained on [14:23] who we are and now all that we do is query  that and then uh Joe legal is my attorney [14:30] at least for the first couple of interactions  with the customer and once it gets to a point [14:34] where we have to do an MSA and all yeah there are  other contractual obligations we then get into uh [14:39] working with an attorney but the point I'm trying  to get to is there are facets of the business [14:45] where you can convenient iently leverage agents.  We have started already using them there. Um [14:52] in terms of uh out sales outreach, we have agents  which are actually calling prospects on our behalf [15:00] explaining what we do. So in a way they are  digital sales agents. So we already have them [15:06] and uh for finance and operations as I said  legal reviews and whatnot uh in terms of uh what [15:15] the outcomes are. So all this is great. Um and  sometimes it becomes an academic exercise when [15:21] you when you go through these slides. Proof  of the pudding is in its taste as they say. [15:27] So the numbers speak for themselves in terms of  what we have been able to achieve internally. [15:31] Right. Um so we we have a 38% uh less design  rework that we have been doing using our agents. [15:40] There is 42% less QA effort that my team does.  there is about 35% less change requests uh [15:46] fewer escape defects all these leading  to the first two that we spoke about [15:54] can I do more with less which is can I  do can I be faster and can I be cheaper [15:59] and the answer to that was 31% faster delivery  and then 25% less run cost right so that's what we [16:05] were able to achieve with this and so it led us to  think can we take this story out to our customers [16:14] and a case study of what we did uh out here  in terms of uh you know working with one of [16:21] our customers. This is AIdriven recall  operations in the automotive industry [16:26] where we are dealing with thousands of dealerships  you know millions of cars related to recalls. [16:34] So how do we essentially handle the entire  recall itself? So we started building agents [16:40] that would help in the recall process. So dealers  would uh you know get a get a car or a vehicle, [16:48] they would go ahead and make the fix based  on the recall. They would take a picture, [16:53] they would upload it and then that picture gets  validated by uh the customer and uh the the [17:00] OEM that we work with and they would in turn go  ahead and validate the claim and whatnot. Right? [17:06] This was a manual process which we used agents to  go ahead and automate right and whenever there is [17:12] a discrepancy it is flagged for human intervention  and the numbers speak for themselves in terms of [17:17] uh you know 73% uh image validation 16 times  turnaround time when compared to manual work [17:26] and then 81% fewer follow-ups many of times  dealers do not submit all the documentation [17:32] we are avoiding that right so that's the kind of  power that we are able to build using AI agents [17:37] and all this is possible at the price point that  we were able to deliver to them and at the time [17:43] that we were able to deliver to them only because  of the approach that I was outlining earlier. [17:48] Um what you're seeing here is a classic dashboard.  I'll leave you with this sort of um image here. [17:58] These are this is a video recording of a live  agent dashboard. So these are agents that are [18:04] doing their work. They are actually capturing  fraud. They are actually going ahead and serving [18:09] IT tickets at an IT desk service  desk. They are doing as HR help desk. [18:16] There is one for concierge and guest  services at a hospitality company. Uh [18:21] there's supply chain intelligence. There are  agents that are deciding trying to figure out [18:25] which part should be shipped from which warehouse  to another warehouse and by what. So all these [18:30] are agents that are already doing this for us in  terms of planning and if you are interested in [18:37] any one of these right uh you can scan the QR code  this will take you to our tokconomics page where [18:43] where you can simply log in and uh try it  out for yourself. The message I have is [18:50] AI has been evolving and there are many companies  trying to you know leverage AI for their own [18:56] benefits making your own business smarter right  we have done that we continue to do that it's an [19:03] it's a journey and we have had tremendous  success with that internally and all that [19:09] we are saying is we could probably help you by  externalizing that uh process for you. Thank you.