Domino Deploys MCP to Automate Regulated Model Training in Cursor

From Vibe Coding to Vibe Modeling: Audit‑Ready GenAI the Model Context Protocol With Domino Data — Ai4

Ai4

Vibe coding fails enterprise compliance because machine learning requires continuous data lineage and dynamic runtime tracking that static code generation cannot deliver.

AI-assisted summary of From Vibe Coding to Vibe Modeling: Audit‑Ready GenAI the Model Context Protocol With Domino Data — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 29, 2025, this presentation features Ankur Sinha, Head of Partner Solutions and GTM at Domino Data Lab. According to the page description, the session explores moving from social-media-hyped 'vibe coding' to enterprise-ready 'vibe modeling,' using the open Model Context Protocol (MCP) and Domino's governed platform with NetApp snapshots to deliver auditable, reliable GenAI.

## TL;DR Ankur Sinha presents 'vibe modeling' as a disciplined evolution of vibe coding, designed to meet the complex, non-linear, and auditable requirements of enterprise data science. Through a live technical demonstration, he shows how the Domino Modeling Assistant pairs Cursor IDE and Claude 3.7 with Domino platform APIs via MCP to automate iterative neural network training, architecture tweaks, and hyperparameter tuning while capturing complete lineage. The presentation also highlights key platform capabilities across data preparation and RAG debugging, bolstered by high-performance NetApp storage integrations. Concluding with a Q&A, Sinha emphasizes why dynamic, retraining-heavy ML lifecycles require dedicated systems of record rather than static code versioning tools.

## Atmosphere & style The presentation carries a pragmatic, polished, and engineer-friendly conference tone that cuts through GenAI hype with real-world enterprise discipline. Sinha delivers a steady, authoritative walkthrough, shifting smoothly between conceptual comparisons and a concrete terminal-and-IDE workflow. It feels like watching an insightful technical briefing where rapid agentic autonomy is carefully anchored by governance, reproducible infrastructure, and live audience interaction.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) Vibe coding enables rapid, natural-language software development in modern IDEs, but is largely optimized for conventional applications and web interfaces. - [01:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=106s) Enterprise machine learning requires moving beyond vibe coding because rapid-fire prompting produces unverified, untraceable outputs that fail regulatory and risk standards. - [01:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=106s) Machine learning lifecycles diverge fundamentally from traditional software engineering due to non-linear experimentation, dynamic data dependencies, and complex multi-artifact deliverables. - [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) Vibe modeling bridges the gap between development speed and scientific rigor by automatically capturing model lineage, parameters, datasets, and execution environments. - [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) The Domino Modeling Assistant leverages the open Model Context Protocol (MCP) to connect IDE agents directly to enterprise compute clusters and governance systems. - [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) In a live demonstration, an MCP-connected Cursor agent running Claude 3.7 autonomously orchestrated remote Domino training jobs on a diabetes classification dataset. - [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) The AI agent iteratively diagnosed overfitting and modified the neural network architecture and optimizer settings directly in code to achieve over 97% accuracy. - [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) The platform supports three core production use cases: automated data preparation, model hyperparameter tuning with MLflow tracking, and RAG application debugging. - [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) Domino integrates with NetApp intelligent data infrastructure, including Amazon FSx for NetApp ONTAP, to deliver snapshot-backed storage with enterprise role-based access control. - [11:40](https://www.youtube.com/watch?v=burzxvDJ22c&t=700s) While Domino remains infrastructure- and storage-agnostic, its partnership with NetApp delivers an optimized native experience for shared enterprise customers. - [11:40](https://www.youtube.com/watch?v=burzxvDJ22c&t=700s) Unlike static software code, machine learning systems require specialized systems of record to manage ongoing model drift, historical lineage, and continuous retraining in production.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) Introduction: Moving from Vibe Coding to Vibe Modeling

Ankur Sinha, Head of Partner Solutions and GTM at Domino Data Lab, welcomes attendees to the session. While the talk is titled *"From Vibe Coding to Vibe Modeling: Audit-Ready GenAI via the Model Context Protocol With Domino Data"*, Sinha clarifies that the core emphasis is on enhancing data science productivity through "vibe modeling." He introduces the presentation roadmap: understanding the fundamentals of vibe coding versus vibe modeling, breaking down how modeling diverges from traditional software engineering concepts, presenting a live demonstration of Domino’s Modeling Assistant, and highlighting the strategic enterprise partnership between Domino Data Lab and NetApp.

Sinha reflects on the emergence of vibe coding as a developer trend. For software engineers who began their careers writing manual code line by line, having the ability to open an Integrated Development Environment (IDE), prompt in natural language, and rapidly translate raw concepts into functional software feels like "a real blessing." However, the current vibe-coding paradigm remains intensely focused on building standard software applications and web interfaces quickly.

```mermaid flowchart LR accTitle: Comparison of Vibe Coding versus Vibe Modeling subgraph VC["Vibe Coding (Software Focus)"] A["Natural Language Prompts"] --> B["Rapid-Fire Code Generation"] B --> C["Flashy Front-End / App Output"] C -.-> D["Risks: Hallucinations, Untraceable State"] end subgraph VM["Vibe Modeling (ML & Data Focus)"] E["Agentic IDE (Cursor / MCP)"] --> F["Automated Compute & Training Jobs"] F --> G["Tracked Artifacts & Lineage"] G --> H["Audit-Ready & Reproducible Models"] end ```

[01:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=106s) The Limits of Vibe Coding in Regulated Machine Learning

While vibe coding produces flashy social media demos and rapid-fire outputs, Sinha emphasizes that its results are often unverified, untraceable, and prone to hallucinations. Enterprise machine learning requires an approach that pairs high velocity with scientific rigor. Vibe modeling bridges this gap by combining speed with trust, enabling data science teams to deliver AI that meets the stringent reproducibility, transparency, and compliance standards demanded by business owners, risk teams, and regulators.

The fundamental reason vibe coding falls short in regulated enterprise environments lies in the structural differences between software development and machine learning lifecycles:

* **Non-linear and Asynchronous Workflows:** Unlike linear software engineering pipelines, model training and experimentation are highly non-linear, experimental, and asynchronous. * **Complex Deliverables:** The final deliverable in data science is not merely an isolated binary or script; it is the entire auditable history, data lineage, hyperparameters, and environment context behind the artifact. * **Requirement for Domain-Specific AI Agents:** General coding assistants know how to complete syntax, but ML workflows require AI agents capable of orchestrating infrastructure, executing remote jobs, and capturing full lineage without trading away developer flexibility.

[03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) Introducing the Domino Modeling Assistant

Sinha introduces the Domino Modeling Assistant, an integration that augments modern IDE coding assistants to execute end-to-end data science tasks—such as iterative experimentation and hyperparameter tuning—while automatically enforcing enterprise governance. It provides full reproducibility, tracking data, code, metrics, and compute artifacts behind the scenes. As Sinha notes, "Coding assistants are definitely increasing the productivity, but data scientists do more than just write code. They build models, and hence modeling assistants are really multiplying the impact for business."

- **Unified Lineage:** Captures model weights, parameters, datasets, and execution environments automatically. - **Enterprise Governance:** Integrates compliance into rapid prototyping without interrupting the data scientist's natural interactive flow. - **Agentic IDE Integration:** Connects local or cloud IDEs directly to scalable enterprise compute clusters via the open Model Context Protocol (MCP).

```mermaid flowchart TD accTitle: Domino Modeling Assistant Architecture subgraph IDE["Developer Environment"] User["Data Scientist"] --> Agent["Cursor IDE Agent (Claude 3.7)"] end subgraph Protocol["Integration Layer"] Agent --> MCP["Domino MCP Server"] end subgraph Backend["Domino Platform & Infrastructure"] MCP --> API["Domino Platform APIs"] API --> Comp["Distributed Compute / Training Jobs"] API --> Reg["System of Record (Lineage, Metadata, MLflow)"] API --> Storage["NetApp Intelligent Data Infrastructure / Volumes"] end ```

[04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) Live Demonstration: Automated Training and Hyperparameter Tuning via MCP

A recorded technical demonstration illustrates how the Domino Modeling Assistant works in practice. The setup pairs the Cursor IDE agent—running Claude 3.7 with reasoning/thinking enabled—with the Domino platform API via an MCP server. The objective is to train and optimize a classification model on a diabetes CSV dataset. The model architecture consists of a configurable four-layer neural network, and the training script accepts standard command-line arguments including epochs, batch size, learning rate, and hidden layer dimensions.

``` +-----------------------------------------------------------------------------------+ | ITERATIVE MODEL OPTIMIZATION DEMO | +---------+------------------------------------+---------------+--------------------+ | Attempt | Architectural & Parameter Changes | Accuracy Rate | Platform Action | +---------+------------------------------------+---------------+--------------------+ | Run 1 | 10 epochs, batch 32, lr 0.0001 | ~96% | Domino Job logged | | Run 2 | More epochs, larger model, lower lr| >96% | Overfitting noted | | Run 3 | Increased dropout rate in PyTorch | Improved | Architecture edit | | Run 4 | Added Adam optimizer weight decay | >97% | Final model chosen | +---------+------------------------------------+---------------+--------------------+ ```

Guided by Cursor system rules, the AI agent automatically reads the local training script, inspects the neural network architecture, validates Domino project settings, and checks Git synchronization between the local IDE and the Domino workspace. It then executes a sequence of automated optimization steps: 1. **Initial Job Execution:** The agent submits a remote training run on Domino configured with 10 epochs, a batch size of 32, small model dimension, and a learning rate of 0.0001, yielding an initial accuracy of 96%. 2. **Experiment Tracking:** The agent monitors job completion, pulls logs, and exposes the specific experiment run in the Domino web UI, ensuring the job is reproducible and accessible to team collaborators. 3. **Overfitting Mitigation:** Noting the performance, the agent initiates a second run with more epochs, a larger model, and a reduced learning rate. Observing risks of overfitting, Claude modifies the neural network architecture directly in code to increase the dropout rate. 4. **Optimizer Adjustment:** On the subsequent iteration, the agent adds weight decay regularization to the Adam optimizer. 5. **Convergence:** The final run reaches over 97% accuracy. The agent presents a structured summary of all iterations and tags the run as a validated candidate model.

[07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) Core Platform Capabilities and the NetApp Partnership

Sinha returns to break down the architectural building blocks of vibe modeling within Domino. The ecosystem relies on three integrated pillars: the IDE/coding assistant running over Domino MCP servers, Domino APIs that orchestrate compute and secure data access, and Domino's centralized system of record that logs all artifacts, metrics, and dependencies.

Beyond classical tabular optimization, Sinha outlines three primary production use cases for the assistant: * **Data Preparation:** Interactive dataset exploration, automated transformation script generation, summary profiling, and programmatic dataset updates. * **Model Tuning:** Automated architectural iterations, dynamic training configuration updates, and unified metric tracking within MLflow. * **RAG Application Debugging:** Automatic retrieval of execution logs, inspection of failure modes, and instant spinning up of isolated diagnostic workspaces.

Sinha highlights the underlying infrastructure partnership with NetApp. The Domino platform integrates with NetApp intelligent data infrastructure, including Amazon FSx for NetApp ONTAP and high-performance NetApp volumes. This integration provides data science teams with scalable, snapshot-backed data access directly through Domino Volumes, pairing elastic GPU/CPU compute orchestration with enterprise Role-Based Access Control (RBAC) and data governance.

[11:40](https://www.youtube.com/watch?v=burzxvDJ22c&t=700s) Q&A: Enterprise Agnosticism and ML vs. Software Lifecycles

During the audience question-and-answer session, attendees ask about platform compatibility, corporate structure, and the fundamental differences between software and model engineering:

- **Storage and Cloud Agnosticism:** Addressing a question regarding hardware lock-in, Sinha clarifies that while Domino is fundamentally infrastructure- and storage-agnostic, the deep strategic partnership with NetApp delivers an optimized, native experience for customers running NetApp storage. He confirms that Domino Data Lab and NetApp remain separate independent companies collaborating on joint solutions. - **Why Model Governance Differs from Code Versioning:** An attendee asks why traditional software engineering tools cannot simply handle data science lifecycles. Sinha explains that software code is largely static once deployed into production, whereas machine learning models are dynamic and intrinsically tied to changing data distributions. "The work starts really after that—the retraining and being able to track all the history," Sinha notes, highlighting why AI development requires specialized systems of record that capture runtime environment state, data drift, and iterative training histories over time.

Sinha invites attendees to review Domino’s open-source MCP server repository on GitHub, read their technical blog posts on vibe modeling, and visit Booth 415 to see live interactive demonstrations of the Domino and NetApp integration.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) From Vibe Coding to Vibe Modeling: Audit-Ready GenAI via the Model Context Protocol With Domino Data — Ankur Sinha

> [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) a real blessing — Ankur Sinha

> [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) Coding assistants are definitely increasing the productivity, but data scientists do more than just write code. They build models, and hence modeling assistants are really multiplying the impact for business. — Ankur Sinha

> [11:40](https://www.youtube.com/watch?v=burzxvDJ22c&t=700s) The work starts really after that—the retraining and being able to track all the history — Ankur Sinha

> building AI that not only works fast but also feels right to business owners, risk teams, and regulators. — Ankur Sinha

> turns hyped speed into audit‑ready reliability. — Ankur Sinha

> Watch an engineer generate a first answer, trace a hallucination through MCP’s chat‑with‑context logs, fix it, trigger scripted validation, and redeploy—all in minutes. — Ankur Sinha

> No theory—just a repeatable checklist to help your own team start vibe modeling tomorrow. — Ankur Sinha

## Mentioned Ankur Sinha, Domino Data Lab, Ai4, NetApp, Model Context Protocol (MCP), Cursor, Claude 3.7, PyTorch, Adam optimizer, MLflow, Amazon FSx for NetApp ONTAP, Domino Volumes, Domino Modeling Assistant, GitHub

## Video description Ankur Sinha, Head of Partner Solutions and GTM, Domino Data Lab

You’ve heard of vibe coding—rapid‑fire prompting that wows on social feeds but wilts under scrutiny. What’s next is vibe modeling: building AI that not only works fast but also feels right to business owners, risk teams, and regulators. In this 20‑minute session we’ll unpack a real‑world regulatory‑assistant built on the open Model Context Protocol (MCP)—and show why running MCP inside Domino’s governed platform (backed by NetApp snapshots) turns hyped speed into audit‑ready reliability. Watch an engineer generate a first answer, trace a hallucination through MCP’s chat‑with‑context logs, fix it, trigger scripted validation, and redeploy—all in minutes. No theory—just a repeatable checklist to help your own team start vibe modeling tomorrow.

[Source](https://www.youtube.com/watch?v=burzxvDJ22c)

中文简报

Domino 在 Cursor 中部署 MCP 实现受监管模型训练自动化 Vibe coding 无法满足企业合规要求，因为机器学习需要持续的数据溯源和动态运行时跟踪，而静态代码生成无法提供这些能力。 Ankur Sinha 提出了作为 vibe coding 规范化演进的“vibe modeling”（氛围建模），旨在满足企业数据科学复杂、非线性及可审计的要求。通过现场技术演示，他展示了 Domino Modeling Assistant 如何通过 MCP 将 Cursor IDE 和 Claude 3.7 与 Domino 平台 API 相结合，在记录完整溯源信息的同时，自动化完成神经网络迭代训练、架构调整与超参数调优。该演示还重点介绍了平台在数据准备和 RAG 调试方面的关键能力，并依托高性能 NetApp 存储集成提供支持。在最后的问答环节中，Sinha 强调了为何动态且高度依赖重新训练的机器学习生命周期需要专门的记录系统，而非仅仅依赖静态代码版本控制工具。 Vibe coding 虽能加速常规软件开发，但在受监管的企业级机器学习场景中极易产生无法追溯且未经审计的输出；而 Vibe modeling 则将快速原型开发与科学严谨性相结合，以满足风控与合规要求。 (1:46) Domino Data Lab 推出 Domino Modeling Assistant，利用开放模型上下文协议（MCP）将现代 IDE 中的 AI 智能体直接接入企业算力与治理系统，自动捕获模型血缘、参数及运行环境。 (3:46) 现场演示展示了在 Cursor 中运行 Claude 3.7 的智能体如何通过 MCP 自主调度 Domino 远程训练作业，并通过迭代修改网络架构与优化器设置，将分类准确率提升至 97% 以上。 (4:47) 该方案覆盖自动数据准备、集成 MLflow 的超参数微调以及 RAG 应用排错三大生产场景，并结合 NetApp 智能数据基础架构实现受快照保护的高性能数据访问与治理。 (7:56)

Reading summary

Domino Data Lab's Ankur Sinha argues that while rapid natural-language code generation accelerates generic software development, regulated enterprise machine learning requires 'vibe modeling'—an approach that binds agentic development environments to automated compute and rigorous lineage tracking via the Model Context Protocol. Limits of Vibe Coding in Regulated Machine Learning (1:46) Sinha notes that standard vibe coding allows software engineers to convert natural-language prompts directly into functional code within an integrated development environment (IDE). However, he contends that the resulting artifacts in enterprise machine learning settings are frequently unverified, untraceable, and prone to hallucinations. Enterprise machine learning requires an approach that pairs development velocity with audit-ready scientific rigor to satisfy business stakeholders, risk teams, and regulatory bodies. According to Sinha, the breakdown occurs because machine learning workflows diverge fundamentally from traditional software engineering. Unlike linear software releases, model development is experimental, non-linear, and asynchronous. The ultimate deliverable is not merely an isolated script or binary, but an auditable history comprising data lineage, hyperparameters, training environments, and evaluation metrics that generic coding assistants cannot independently manage. Architecture of the Domino Modeling Assistant (3:46) To adapt conversational AI assistants to data science, Domino introduced the Domino Modeling Assistant. The framework connects local or cloud-based agentic IDEs to Domino's scalable compute and centralized system of record using the open Model Context Protocol (MCP). This integration executes tasks such as automated experimentation and hyperparameter tuning while enforcing governance behind the scenes. Sinha highlights three core production use cases enabled by this architecture: interactive data preparation and transformation, automated model architecture tuning tracked through MLflow, and automated retrieval-augmented generation (RAG) debugging via isolated diagnostic workspaces. Behind these workflows, the platform logs model weights, parameters, datasets, and runtime environments to preserve full reproducibility without interrupting the practitioner's interactive flow. Automated Hyperparameter Optimization in Practice (4:47) A technical demonstration illustrated the assistant's mechanics by pairing a Cursor IDE agent—running Claude 3.7 with extended thinking—with Domino APIs via an MCP server. Tasked with training and optimizing a four-layer classification neural network on a tabular diabetes dataset, the agent read local scripts, inspected network configurations, and validated Git synchronization between the IDE and Domino workspaces. The AI agent autonomously coordinated four iterative training cycles on Domino compute. An initial baseline job (10 epochs, batch size 32, learning rate 0.0001) achieved approximately 96% accuracy. To counter observed overfitting risks in subsequent runs, Claude directly modified the PyTorch code to raise dropout rates and introduced weight decay regularization to the Adam optimizer, ultimately pushing accuracy above 97% while automatically logging all runs to Domino's web interface. Lifecycle Governance and Infrastructure Integration (7:56) Addressing post-deployment requirements, Sinha emphasizes that model governance cannot rely solely on standard software version control. Unlike deployed software code, which remains relatively static, machine learning models degrade as live data distributions shift, necessitating continuous tracking of retraining history, environment states, and data drift over time. To support these compute- and data-intensive lifecycles, Sinha outlines Domino's technical partnership with NetApp. While clarifying that Domino remains cloud- and storage-agnostic, he notes that integrations with NetApp intelligent data infrastructure—such as Amazon FSx for NetApp ONTAP—grant teams snapshot-backed volume access, enterprise role-based access control, and elastic compute orchestration directly within the governance system of record. Domino frames vibe modeling as the necessary bridge between conversational AI velocity and enterprise compliance; however, sustained success relies on whether autonomous agent-led code modifications can consistently maintain transparent traceability as underlying data distributions and infrastructure scale.

精炼阅读摘要

Domino Data Lab 合作伙伴解决方案兼市场拓展负责人 Ankur Sinha 指出，当前基于自然语言的“氛围编码”（Vibe Coding）虽能快速构建软件，但在受监管的企业环境中存在不可追溯和幻觉风险；企业必须依托模型上下文协议（MCP）将智能 IDE 与企业级记录系统打通，转向兼顾敏捷性与审计合规的“氛围建模”（Vibe Modeling）。 软件与机器学习生命周期的本质分歧 (1:46) Sinha 强调，以自然语言生成代码为特征的氛围编码虽然在前端与常规应用开发中大幅提升了开发速度，但其输出往往未经严格验证且状态不可追踪，极易引入幻觉。在金融、医疗等强监管领域，机器学习工作流具有非线性和高度异步实验的特征，这与传统线性软件工程流水线存在根本差异。在数据科学中，最终交付物并非单一的可执行脚本或二进制程序，而是包含模型权重、超参数、数据集血缘及运行时环境上下文的完整可审计历史。 因此，企业级人工智能开发不能仅依赖通用的语法补全辅助工具。通用代码助手虽然能够加速编写代码，但无法管理底层算力、远程调度训练作业以及捕获完整的合规血缘。氛围建模的核心目标正是在保留开发者交互灵活性的同时引入科学严密性，确保模型输出能够满足业务负责人、风险管理团队以及监管机构对透明度和重现性的严格要求。 基于 MCP 的建模助手架构与治理 (3:46) 为了弥合快速原型构建与企业级治理之间的鸿沟，Domino 推出了基于模型上下文协议（Model Context Protocol, MCP）的建模助手（Domino Modeling Assistant）。该方案构建了三层整合架构：顶层为数据科学家的交互式开发环境（如运行 Claude 3.7 并开启推理思考能力的 Cursor IDE Agent），中间层为开源的 Domino MCP 服务器，底层直连 Domino 平台 API 及其记录系统（System of Record）。 通过开放的 MCP 协议，建模助手能够将本地或云端 IDE 直接挂接到企业可扩展的计算集群上，自动在后台记录代码变更、超参数配置、度量指标与计算产物，并将实验元数据同步至 MLflow。Sinha 指出，数据科学家不仅需要编写代码，更核心的任务是构建模型；该架构使合规审计在不打断开发者自然交互流程的前提下静默完成，实现了数据、代码与环境的全面统一溯源。 自动化模型微调与参数迭代实践 (4:47) 在现场技术演示中，系统展示了 Cursor 智能体通过 Domino MCP 服务器优化糖尿病分类模型的完整过程。该模型基于四层神经网络，训练脚本支持通过命令行参数配置训练轮数（Epochs）、批大小（Batch Size）、学习率（Learning Rate）以及隐藏层维度。在系统规则引导下，智能体首先自主读取本地训练脚本并验证 Git 同步与项目设置，随后提交了首个远程训练作业（10 轮迭代、批大小 32、学习率 0.0001），初步实现了约 96% 的准确率，相关日志自动记录至 Domino 平台。 面对后续迭代中出现的过拟合现象，智能体展现出端到端的架构干预能力：它不仅增加了迭代轮数并调低学习率，还直接在代码中修改 PyTorch 网络结构以提高 Dropout 比率；在第四次迭代中，智能体进一步在 Adam 优化器中引入权重衰减（Weight Decay）正则化，最终将模型准确率提升至 97% 以上。整个优化链条中的各次尝试、参数变化与性能对比均被整理成结构化摘要，并作为验证候选模型留存。 企业级存储协同与核心生产场景 (7:56) Sinha 将 Domino 氛围建模的生产落地场景划分为三大核心方向：其一是交互式数据探索与自动化转换脚本生成的数据准备；其二是包含动态架构迭代与指标追踪的模型微调；其三是自动抓取执行日志并一键启动隔离诊断工作区的检索增强生成（RAG）应用调试。在底层基础设施层面，Domino 深度整合了 NetApp 智能数据架构（包括 Amazon FSx for NetApp ONTAP 及高性能 NetApp 卷）。 该合作通过 Domino 卷直接为数据科学团队提供基于快照备份的可扩展数据访问，同时将弹性 GPU/CPU 算力编排与企业级基于角色的访问控制（RBAC）深度绑定。在问答环节中，Sinha 特别澄清，尽管与 NetApp 保持独立公司的战略协作并针对其存储进行优化，Domino 平台本身在架构上保持对底层基础设施和云存储的完全中立。 模型动态演化与代码版本控制的区别 (11:40) 针对观众提出的“为何传统软件版本控制工具无法直接满足数据科学需求”的疑问，Sinha 作出理论辨析。他指出，传统软件代码一旦编译并部署上线，其内在逻辑通常保持静态稳定；而机器学习模型在上线后与外部数据分布紧密绑定，其实际表现会随现实环境的数据漂移（Data Drift）而持续演变。 Sinha 表示，“真正的工作往往在模型部署之后才刚刚开始”，这包括持续的模型重训以及对全历史演变轨迹的追踪。单纯依赖代码版本管理系统（如 Git）无法捕捉动态的运行时环境状态、训练数据快照以及性能指标漂移，企业必须建立专门的系统级记录体系，才能保证模型在整个生命周期内的可追溯性与持续合规。 Sinha 总结认为，企业应通过开放协议实现从孤立代码生成到全流程受控建模的转变，以兼顾研发速度与合规韧性。然而，本次演讲主要依赖受控环境下的结构化表格分类任务进行演示，在面对更复杂的非结构化多模态大模型微调或异构跨云分布式作业时，该架构的调试开销与实际容错能力仍有待更多生产实践检验。

Detailed notes — zh

## Background 该内容由行业会议组织方 Ai4 于 2025 年 8 月 29 日在 YouTube 平台发布，由 Domino Data Lab 合作伙伴解决方案与 GTM 负责人 Ankur Sinha 带来主题分享。视频简介指出，演讲聚焦于如何超越社交网络上盛行却难以经受合规审视的“氛围编程（Vibe Coding）”，转而利用开源的模型上下文协议（MCP）与 Domino 受控平台构建审计就绪的“氛围建模（Vibe Modeling）”工作流。 ## TL;DR 演讲深入剖析了数据科学由于其非线性探索与严格审计追溯要求，无法直接套用传统软件工程 Vibe Coding 范式的根本原因，进而提出了兼具敏捷迭代与企业级治理的 Vibe Modeling 理念。现场重点展示了基于 Cursor IDE、Claude 3.7 Thinking 与 Domino MCP Server 的自动化建模助手，演示了从代码自检、异步实验触发到过拟合自主诊断与超参数调优的全流程闭环。演讲还详细解析了覆盖数据清洗、模型微调与 RAG 排错的底层架构体系，并介绍了与 NetApp 高性能存储设施的原生集成方案。最后在问答中进一步明确了模型部署后应对数据动态漂移与持续再训练对于全生命周期治理的核心价值。 ## Atmosphere & style 整场分享节奏明快、务实严谨，充满浓厚的工业界落地导向与工程师对话氛围。演讲者以极具亲和力且条理清晰的语调，迅速将观众从浮躁的“快餐式提示词生成”拉回到对企业合规与严密复现的理性思考中。当现场演示中 Claude 3.7 Thinking 驱动 Domino MCP 自动诊断过拟合并修改代码时，现场展现出前沿 AI Agent 协作的高效与惊艳感，互动环节则进一步体现了对模型动态生命周期管理的深刻技术洞察。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) 本次分享的核心在于从强调快速生成代码的 Vibe Coding 迈向兼顾速度与治理的 Vibe Modeling（氛围建模），旨在大幅提升数据科学团队的生产力与合规水平。 - [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) 传统的 Vibe Coding 主要服务于线性软件工程，虽能快速拼装应用，但存在幻觉高发、不可溯源和缺乏合规审计能力的缺陷，难以直接应用于严格监管场景。 - [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) 数据建模本质上是非线性且高度探索性的过程，机器学习项目的交付物不仅是最终模型权重，更包含推导过程中的全部历史实验记录与上下文血缘。 - [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) Domino Modeling Assistant 的定位并非替代现有代码编辑器，而是通过增强现有的 IDE 编程助手，赋予其驱动端到端数据科学异步实验与超参数调优的能力。 - [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) 基于模型上下文协议（MCP），Cursor Agent 结合 Claude 3.7 Thinking 能够无缝对接 Domino 平台 API，自动化执行代码同步、远程任务触发与指标轮询。 - [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) 在糖尿病预测模型演示中，Agent 能自主识别过拟合风险并重构代码（提高 Dropout 比率及引入权重衰减），最终将准确率优化至 97% 以上并生成完整复盘报告。 - [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) Domino 梳理了 Vibe Modeling 的三大核心落地场景：数据探索与清洗转换、自动化模型调优与 MLflow 追踪、以及 RAG 应用排错与隔离工作空间创建。 - [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) 系统具备 IDE 无关性（IDE Agnostic），并通过集中的核心记录系统（System of Record）统一管理模型资产、环境快照及血缘关系，确保实验完全可复现。 - [10:33](https://www.youtube.com/watch?v=burzxvDJ22c&t=633s) 在基础设施层，Domino 与 NetApp 深度集成（支持 AWS 上的 FSx for NetApp ONTAP），结合平台 RBAC 访问控制为建模提供安全、弹性的底层存储卷支持。 - [10:33](https://www.youtube.com/watch?v=burzxvDJ22c&t=633s) Domino 已将官方的 Domino MCP Server 在 GitHub 开源，帮助开发者社区快速构建审计就绪的生成式 AI 建模工作流。 - [11:49](https://www.youtube.com/watch?v=burzxvDJ22c&t=709s) 软件工程在部署后代码相对静态，而机器学习模型时刻面临数据漂移；模型生命周期中最关键的挑战始于上线后的持续再训练、历史追溯与上下文治理。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=burzxvDJ22c&t=5s) 演讲开场与核心主题：从“氛围编程”迈向“氛围建模” Domino Data Lab 合作伙伴解决方案与 GTM（Go-to-Market）负责人 Ankur Sinha 首先向参会者致谢，并正式引出当天的演讲主题。虽然演讲的最初预定题目是《从 Vibe Coding 到 Vibe Modeling：审计就绪的生成式 AI》，但在梳理演示内容的过程中，团队意识到这场分享更深层的核心在于**如何通过“氛围建模”（Vibe Modeling）大幅提升数据科学团队的生产力**。 演讲将围绕四大模块逐步展开： 1. 厘清何为 Vibe Coding，以及它与 Vibe Modeling 之间的本质差别； 2. 深入探讨为什么数据建模从底层逻辑上不同于传统的软件工程概念； 3. 现场演示 Domino 团队基于模型上下文协议（MCP，Model Context Protocol）构建的“建模助手”（Modeling Assistant）； 4. 介绍 Domino 与本次联合赞助方 NetApp 的战略合作价值。 --- ### [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) Vibe Coding 与 Vibe Modeling 的本质差异与企业诉求 对于经历过传统软件开发时代的工程师而言，如今能够直接在 IDE 中使用自然语言对话、快速将创意与概念转化为现实代码，无疑是一种“巨大的恩赐”（"a real blessing"）。然而，现有的 Vibe Coding 主要服务于传统软件工程，核心诉求在于如何光速把应用程序拼装上线。这种快节奏的连环 Prompting 固然能在社交网络上带来炫目惊艳的效果，但其产出往往伴随着不可靠、不可溯源以及模型幻觉等严重隐患。 相比之下，Vibe Modeling 的目标是在保留极致速度的同时，注入严谨性（rigor）与可复现性（reproducibility），以满足业务负责人与合规监管团队的严格要求。Vibe Modeling 从一开始就将结构化调试、全链路日志记录与透明可审计性深度嵌入体系之中，真正做到将**“速度与信任打包融合在单一平台中”**（"having speed plus trust baked into one single area and one single platform"）。 ```mermaid flowchart LR accTitle: Vibe Coding 与 Vibe Modeling 模式对比 subgraph VC["Vibe Coding（氛围编程）"] A["连环 Prompt 交互"] --> B["快速生成代码/应用"] B --> C["痛点：幻觉频出/不可溯源/缺乏审计"] end subgraph VM["Vibe Modeling（氛围建模）"] D["自然语言意图引导"] --> E["自动化异步实验编排"] E --> F["结果：全生命周期血缘/可审计/完全复现"] end ``` 在受到严格监管的企业环境中，数据科学与传统软件工程存在着两条根本性的分水岭： - **过程特性完全不同**：传统软件开发生命周期相对线性，而数据科学建模本质上是非线性的、异步的探索过程； - **交付物定义不同**：机器学习项目的交付物绝不仅仅是最终跑出来的模型权重文件，更包含达成该模型背后的**全部历史实验记录与上下文背景**。如果无法随时回溯历史、证明某个指标是如何一步步推导出来的，模型就无法通过合规审计。 因此，企业亟需一种专为机器学习场景量身定制的 AI Agent——它能够在不牺牲灵活性与敏捷度的前提下，以结构化、规范化的方式参与到模型的全生命周期中。 --- ### [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) 方案发布：Domino Modeling Assistant 的定位与核心理念 为了填补这一空白，Domino 推出了 Domino Modeling Assistant。该方案并没有另起炉灶去替代开发者喜爱的代码编辑器，而是直接增强（augment）开发者现有的 IDE 编程助手，赋予其驱动端到端数据科学工作流的能力，覆盖迭代式实验探索、超参数调优等复杂任务。 最关键的是，在助手自主执行实验的过程中，底层的治理机制与合规审计能力会自动生效，完整保留历史轨迹与血缘关系（lineage）。编程助手固然提升了敲代码的效率，但数据科学家的核心职责是**构建与验证模型**，而建模助手才是真正成倍放大业务价值的核心引擎。 - **赋能目标**：将自然语言驱动的 Agent 能力直接注入数据科学全流程； - **合规兜底**：所有异步任务在受控环境中执行，保留完整参数、环境快照及执行日志； - **价值转变**：从单一的代码辅助编写（Code Completion）升级为全局的实验与调优编排（Model Orchestration）。 --- ### [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) 实战演示：基于 Cursor 与 MCP 的端到端糖尿病预测模型优化 现场进入实操演示环节：演示者展示了 Cursor Agent 如何通过 Model Context Protocol（MCP）服务器无缝对接 Domino 平台的 API，完成一个完整的分类模型训练与自动调优用例。任务目标是基于一份简单的 CSV 特征数据集，训练一个用于预测糖尿病的神经网络模型。 该模型采用可灵活配置隐藏层维度的 4 层全连接神经网络架构。训练脚本封装了标准的训练循环，支持通过命令行传入训练轮数（Epochs）、批大小（Batch Size）、学习率（Learning Rate）以及模型尺寸等超参数。 演示者仅在 Cursor 对话框中输入了一条极简的 Prompt，要求模型理解现有的训练脚本，并通过多次尝试自动完成调优。在后台预设的 Cursor Rules 指引下，底层搭载的 **Claude 3.7 Thinking** 认知模型自动调用 Domino MCP 提供的各项工具，展开了一场丝滑的自动化迭代： ```mermaid flowchart TD accTitle: Claude 3.7 Thinking 与 Domino 自动调优闭环 A["解析训练脚本与工程配置"] --> B["验证 IDE 与 Domino 的 Git 状态同步"] B --> C["Run 1: 基础超参探索 (Acc: 96%)"] C --> D["Run 2: 扩大模型容量 (准确率提升)"] D --> E["发现过拟合风险，自动修改代码提高 Dropout 比例"] E --> F["Run 3: 进一步为 Adam 优化器引入 Weight Decay"] F --> G["Run 4: 达成最优候选模型 (Acc: >97%) 并输出完整复盘报告"] ``` 整个调优过程展现出极高的智能度与严谨性： 1. **环境自检与首轮实验**：Claude 3.7 Thinking 首先阅读了训练脚本、模型架构与 Domino 项目配置，主动检查并确保本地 IDE 与 Domino 远端的 Git 代码处于同步状态。随后触发首个 Domino 远程 Job（配置为：10 Epochs、Batch Size 32、Small 模型架构、学习率 0.001）。 2. **状态轮询与结果回传**：在确认任务执行完毕后，Agent 自动拉取指标，确认第一轮实验取得了 96% 的基准准确率。同时，Agent 自动在浏览器中打开该次实验在 Domino 界面中的专属运行记录，所有计算依赖与参数均被完整记录，团队协作者可一键复现。 3. **自主诊断与代码级重构**：Agent 自主决定发起第二轮迭代，尝试增加训练轮数、增大网络尺寸并微调降低学习率，成功获得了更高的准确度。但更亮眼的是，Agent 主动警惕到了过拟合风险（"Cursor has decided it's worried about overfitting"），并自主提出修改模型架构代码，适当调高 Dropout 比率。 4. **终极优化与总结报告**：在检查第三次运行结果后，Agent 又在代码中为 Adam 优化器追加了权重衰减（Weight Decay）逻辑。在最后一次 Domino Job 执行结束后，Agent 自动生成了一份漂亮的调优总结，明确指出最终模型准确率已超过 97%，并将其标记为达标候选模型。 整个过程中所有的代码修改、参数调优与运行时状态，全部被 Domino 平台精准捕获并实现审计可溯源，为数据科学工作流提供了一种既安全又极具可维护性的 Agent 协作新范式。 --- ### [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) 技术架构、三大典型场景与平台核心价值 在演示结束后，演讲者深入剖析了支撑这套 Vibe Modeling 体系的底层技术架构。其核心是由 IDE 前端、协议通信层、算力编排层与核心存证系统共同构成的协同网络： ```mermaid flowchart TD accTitle: Domino Modeling Assistant 技术架构图 subgraph IDE_Layer["开发者交互层"] IDE["IDE 环境 (如 Cursor)"] Agent["智能编码助手 (Claude 3.7 Thinking)"] IDE --- Agent end subgraph Protocol_Layer["协议与接口层"] MCP["Domino MCP Server"] API["Domino Platform APIs"] end subgraph Engine_Layer["受控算力与存证中心"] Compute["算力自动化调度 / 数据访问 / 任务编排"] SOR["Domino 核心记录系统 (System of Record)"] SOR_Details["实验追踪 / 指标比对 / 模型注册表 / 资产血缘"] SOR --- SOR_Details end IDE_Layer -->|"MCP 协议交互"| MCP MCP -->|"API 调用"| API API -->|"触发与管理"| Compute Compute -->|"写入产物与快照"| SOR ``` 基于这套架构，Domino 梳理并落地了三大核心落地应用场景： - **数据准备与清洗（Data Preparation）**：自主探索新数据集、自动编写特征转换脚本、总结分布规律并一键应用数据变更； - **模型调优与训练（Model Tuning）**：自动化探索与重构模型网络架构、测试多组超参数配置，并将全部运行指标无缝同步至 MLflow 进行可视化追踪； - **RAG 应用排错（Debugging a RAG App）**：自动抓取调用链日志、定位召回或生成失败的根因，并为工程师自动创建用于现场排错的隔离工作空间（Workspace）。 演讲者强调，上述三大场景仅仅是一个开端，更多进阶功能已列入产品路线图。Domino Modeling Assistant 带来的核心优势在于：不仅具备 IDE 无关性（IDE Agnostic），让开发者自由选择喜爱的开发界面；更提供了一个集中的记录系统（Central System of Record），统一管理模型资产、血缘关系与环境依赖，使企业在享受生成式 AI 爆发式生产力的同时，守住合规、安全与可审计的底线。 ### [10:33](https://www.youtube.com/watch?v=burzxvDJ22c&t=633s) NetApp 底层存储支持与 Vibe Modeling 资源生态 在平台的治理与合规层面，Domino 提供了基于角色的安全访问控制（RBAC），确保所有操作都在受控且合规的安全平台内运行。而在底层存储方面，平台依托于 NetApp 提供的智能数据基础设施作为强力支撑。目前系统已原生支持 AWS 上的 FSx for NetApp ONTAP（FSx bundle），并结合了平台自身的数据集（Datasets）管理功能。通过与 NetApp 底层卷存储（NetApp Volumes）的深度集成，用户能够通过 Domino Volumes 解锁极具弹性、高度可扩展且安全的数据访问能力。 演讲者随后向现场参会者分享了相关的参考资料与开源生态资源。本次演讲的核心观点与技术框架源自一篇深度的 LinkedIn 博客文章，官方同时还发布了专门探讨 Vibe Modeling 的系列博客供进一步阅读。为了促进社区实践与快速落地，Domino 已将官方的 Domino MCP Server（Model Context Protocol 服务端）开源并托管在 GitHub 公开仓库中。此外，参会者可以通过合作伙伴页面或直接前往 415 号展位，现场观看完整的平台端到端演示。 ```mermaid flowchart TD accTitle: Domino and NetApp Integrated Architecture subgraph DominoPlatform["Domino 治理与计算层"] A["企业级 RBAC 安全管控"] B["Domino MCP Server<br/>(GitHub 开源)"] C["Domino Datasets / Volumes 管理"] end subgraph StorageLayer["NetApp 智能数据基础设施"] D["Amazon FSx for NetApp ONTAP"] E["NetApp 高性能底层存储卷"] end A --- C B --- C C --> D D --> E ``` - **开源与文档资源**：Domino MCP Server 已在 GitHub 完全开源公开；相关技术博客涵盖了从 Vibe Coding 到 Vibe Modeling 的演进逻辑。 - **展位与演示信息**：官方团队驻扎在大会 415 号展台，提供包含 MCP 调试、治理日志追溯与自动化验证的全流程实时演示。 --- ### [11:49](https://www.youtube.com/watch?v=burzxvDJ22c&t=709s) 现场问答：生态兼容性与模型全生命周期管理的本质差异 在随后的问答环节中，有参会者询问了关于基础设施绑定及公司关系的疑问。演讲者明确指出，Domino 在技术底层始终保持技术中立与架构无关（Agnostic）的特性，可以灵活适配多种底层环境；但通过与 NetApp 的深度合作，双方联合打造了开箱即用的原生集成体验，让用户可以直接在 Domino 内部无缝调用 NetApp 的高级存储能力。针对是否隶属于 NetApp 的提问，演讲者澄清 Domino Data Lab 是一家独立运营的软件公司，与 NetApp 保持着紧密的战略合作伙伴关系。 针对建模（Modeling）与传统软件编码（Coding）之间区别的提问，演讲者深入剖析了数据科学与传统软件工程在生命周期上的根本差异。软件工程在代码部署上线后通常是静态不变的，除非发布新版本，否则运行逻辑不会自行演化；但数据科学与机器学习则截然不同——数据和模型时刻都在发生动态漂移。因此，模型生命周期中“真正繁重的工作往往始于部署到生产环境之后”，包括持续的模型再训练（Retraining）、数据与模型版本的全量历史追溯（History Tracking）以及上下文审计。Vibe Modeling 及 Domino 平台正是专门针对模型特有的动态演进需求而构建，而非单纯服务于传统的静态软件应用。 - **基础设施中立性**：Domino 保持多云与存储中立，但为 NetApp 存储架构提供了深度优化的原生集成支持。 - **软件与模型生命周期的核心区别**： - **传统软件**：部署后代码处于静态（Static code），运维重点在服务可用性。 - **数据与 AI 模型**：上线后数据分布持续变化，必须依赖完善的历史回溯、再训练机制与上下文治理。 ## Notable quotes > [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) 对于经历过传统软件开发时代的工程师而言，如今能够直接在 IDE 中使用自然语言对话、快速将创意与概念转化为现实代码，无疑是一种“巨大的恩赐”（a real blessing）。 — Ankur Sinha > [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) having speed plus trust baked into one single area and one single platform — Ankur Sinha > [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) 传统软件开发生命周期相对线性，而数据科学建模本质上是非线性的、异步的探索过程。 — Ankur Sinha > [01:17](https://www.youtube.com/watch?v=burzxvDJ22c&t=77s) 机器学习项目的交付物绝不仅仅是最终跑出来的模型权重文件，更包含达成该模型背后的全部历史实验记录与上下文背景。 — Ankur Sinha > [03:46](https://www.youtube.com/watch?v=burzxvDJ22c&t=226s) 编程助手固然提升了敲代码的效率，但数据科学家的核心职责是构建与验证模型，而建模助手才是真正成倍放大业务价值的核心引擎。 — Ankur Sinha > [04:47](https://www.youtube.com/watch?v=burzxvDJ22c&t=287s) Cursor has decided it's worried about overfitting — Ankur Sinha > [07:56](https://www.youtube.com/watch?v=burzxvDJ22c&t=476s) 不仅具备 IDE 无关性（IDE Agnostic），让开发者自由选择喜爱的开发界面；更提供了一个集中的记录系统（Central System of Record），统一管理模型资产、血缘关系与环境依赖。 — Ankur Sinha > [11:49](https://www.youtube.com/watch?v=burzxvDJ22c&t=709s) 真正繁重的工作往往始于部署到生产环境之后。 — Ankur Sinha > [11:49](https://www.youtube.com/watch?v=burzxvDJ22c&t=709s) 传统软件在部署后代码处于静态，而数据与 AI 模型在上线后数据分布持续变化，必须依赖完善的历史回溯、再训练机制与上下文治理。 — Ankur Sinha ## Mentioned Domino Data Lab, Ankur Sinha, NetApp, Model Context Protocol (MCP), Domino Modeling Assistant, Cursor, Claude 3.7 Thinking, Amazon FSx for NetApp ONTAP, MLflow, GitHub, LinkedIn, AWS ## Video description Ankur Sinha, Head of Partner Solutions and GTM, Domino Data Lab You’ve heard of vibe coding—rapid‑fire prompting that wows on social feeds but wilts under scrutiny. What’s next is vibe modeling: building AI that not only works fast but also feels right to business owners, risk teams, and regulators. In this 20‑minute session we’ll unpack a real‑world regulatory‑assistant built on the open Model Context Protocol (MCP)—and show why running MCP inside Domino’s governed platform (backed by NetApp snapshots) turns hyped speed into audit‑ready reliability. Watch an engineer generate a first answer, trace a hallucination through MCP’s chat‑with‑context logs, fix it, trigger scripted validation, and redeploy—all in minutes. No theory—just a repeatable checklist to help your own team start vibe modeling tomorrow. [Source](https://www.youtube.com/watch?v=burzxvDJ22c)

Source quotes

[1:17] Ankur Sinha: 对于经历过传统软件开发时代的工程师而言，如今能够直接在 IDE 中使用自然语言对话、快速将创意与概念转化为现实代码，无疑是一种“巨大的恩赐”（a real blessing）。 [1:17] Ankur Sinha: having speed plus trust baked into one single area and one single platform [1:17] Ankur Sinha: 传统软件开发生命周期相对线性，而数据科学建模本质上是非线性的、异步的探索过程。 [1:17] Ankur Sinha: 机器学习项目的交付物绝不仅仅是最终跑出来的模型权重文件，更包含达成该模型背后的全部历史实验记录与上下文背景。 [3:46] Ankur Sinha: 编程助手固然提升了敲代码的效率，但数据科学家的核心职责是构建与验证模型，而建模助手才是真正成倍放大业务价值的核心引擎。 [4:47] Ankur Sinha: Cursor has decided it's worried about overfitting [7:56] Ankur Sinha: 不仅具备 IDE 无关性（IDE Agnostic），让开发者自由选择喜爱的开发界面；更提供了一个集中的记录系统（Central System of Record），统一管理模型资产、血缘关系与环境依赖。 [11:49] Ankur Sinha: 真正繁重的工作往往始于部署到生产环境之后。 [11:49] Ankur Sinha: 传统软件在部署后代码处于静态，而数据与 AI 模型在上线后数据分布持续变化，必须依赖完善的历史回溯、再训练机制与上下文治理。

Source quotes

[0:05] Ankur Sinha: From Vibe Coding to Vibe Modeling: Audit-Ready GenAI via the Model Context Protocol With Domino Data [0:05] Ankur Sinha: a real blessing [3:46] Ankur Sinha: Coding assistants are definitely increasing the productivity, but data scientists do more than just write code. They build models, and hence modeling assistants are really multiplying the impact for business. [11:40] Ankur Sinha: The work starts really after that—the retraining and being able to track all the history Ankur Sinha: building AI that not only works fast but also feels right to business owners, risk teams, and regulators. Ankur Sinha: turns hyped speed into audit‑ready reliability. Ankur Sinha: Watch an engineer generate a first answer, trace a hallucination through MCP’s chat‑with‑context logs, fix it, trigger scripted validation, and redeploy—all in minutes. Ankur Sinha: No theory—just a repeatable checklist to help your own team start vibe modeling tomorrow.

Original transcript

[0:05] So uh good morning everybody uh thank [0:08] you for attending the session today. So [0:10] uh what we are going to be talking about [0:12] today is uh W modeling and w coding and [0:16] u I mean the title all the title on the [0:20] thing says that it's going to be from [0:22] why modeling w coding to w modeling [0:24] audit ready genai but I think as I was [0:27] putting this presentation together what [0:28] we thought is it is more focused on how [0:30] we enhance the data science productivity [0:33] uh with with modeling so that's what we [0:35] are going to be focusing on today uh [0:37] just to quickly introduce myself I'm an [0:39] I work with Domino Data Lab. I run [0:41] partnerships for Domino. And uh what [0:44] we're going to be looking at today is uh [0:50] one let's let's start with understanding [0:52] what WIP coding is and how is it [0:54] different uh from WI modeling. Then [0:56] we'll dive deeper into why modeling is [0:59] different from really the software [1:00] engineering core concepts. uh I have a [1:03] small demo for all of you uh of the [1:04] modeling assistant that we have built [1:06] and then a little bit of a plug about [1:08] Domino plus NetApp because we have we [1:10] are sponsoring this with NetApp and how [1:12] we are offering a strategic partnership [1:14] there. [1:17] So to talk about W coding uh so how many [1:20] of us have started as a software [1:22] engineer back in the day start with [1:24] coding I think this is this is a real [1:25] blessing being able to just get into an [1:28] IDE and start talking in natural [1:30] language and start quickly coding and [1:32] bringing ideas to concepts to and ideas [1:34] into into reality. Uh but one of the [1:37] things that we have noticed really [1:39] heavily focused on software engineering [1:41] and everything is really about how we [1:43] can quickly bring apps into into life uh [1:46] uh and all about that. With VIP coding [1:49] you definitely get rapid fire prompting [1:51] you definitely get flashy results right [1:53] uh but sometimes the outputs are [1:56] unreliable untraceable and hallucinated [1:58] too. Uh so over there what what we what [2:02] we look at in W modeling is what we are [2:05] focusing on is we are combining we're [2:06] trying to combine the speed with rigor [2:08] and reproducible AI that is your your [2:11] needs for your business users and [2:13] stakeholders uh definitely transparent [2:16] and auditable uh for regulatory and [2:18] compliance purposes and include [2:20] structured debugging logging and [2:21] reproducibility from the start. So [2:23] basically if you have to think about Y [2:25] modeling you think about having speed [2:27] plus trust baked into one single area [2:29] and one single platform [2:31] here and when when we think about W [2:34] coding why is it what we have noticed [2:36] and why why we think it isn't enough for [2:38] enterprise use cases is because teams [2:41] generally would love rapid [2:42] experimentation with generative web but [2:45] I think one of the core things to really [2:46] focus on is when when you're doing rapid [2:48] experimentation you also uh when you're [2:51] doing it in regulated enterprises you [2:53] want to have the history and the context [2:55] behind it. You want to have reproducible [2:57] results that that's coming out of it. So [2:59] two key differentiator that we have seen [3:02] specifically in a model life cycle [3:04] versus a software engineering life cycle [3:06] is that uh software engineering uh [3:09] basically modeling is technically [3:10] nonlinear. They're pretty the [3:12] experiments are asynchronous. The [3:14] deliverable technically needs to be [3:15] audited because you should be able to go [3:17] back in time look at your results as to [3:19] how you're producing it, what you're [3:20] producing there and and the deliverable [3:22] is not just the model. It also the [3:24] history of that work that's that that [3:26] you have created. Uh so basically the [3:29] need that we have analyzed after working [3:32] on this is that there is there needs to [3:34] be AI agents which is ready for ML and [3:36] need to part participate in the entire [3:38] life cycle. So basically a fast [3:40] structured approach without sac [3:42] sacrificing speed for flexibility. [3:46] Hence I would like to introduce uh [3:49] Domino modeling assistant here. What it [3:52] gives you is it is giving you it [3:54] augments your ID coding assistant to [3:56] execute end toend data science workflows [3:58] such as iterative experimentation and [4:00] model tuning but it is doing while [4:03] supporting uh governance and compliance [4:05] requirements. So it's fully reproducible [4:07] and uh it maintains the history and [4:09] lineage behind it. So that's that's like [4:11] the core of of any model life cycle that [4:14] you would think about going back in time [4:16] having the full full reproducibility [4:17] full governance baked in. That is what [4:19] what you pro uh what we provide. Now [4:22] coding assistants are definitely [4:24] increasing the productivity but data [4:25] scientists do more than just write code. [4:27] Uh they build models and hence modeling [4:30] assistants are really multiplying the [4:31] impact for business. With this, I will [4:34] quickly take you through a small demo. [4:47] >> I'd like to show you a quick model [4:48] training and optimization use case using [4:50] the Domino modeling assistant. We'll do [4:53] this with cursors agent connected to [4:55] Domino's platform API using our MCP [4:58] server. We're going to work with a [5:00] simple CSV data set to train our [5:02] classification model to predict diabetes [5:04] with a simple set of features. [5:07] Here you can see the diabetes model [5:09] architecture, a simple four-layer neural [5:12] net with hidden dimension sizes that are [5:13] configurable. [5:16] Our training script defines a typical [5:18] training loop for this model and runs [5:20] from the command line with arguments for [5:22] setting typical parameters like number [5:24] of epochs, batch size, learning rate, [5:26] and the model size. I'm going to use a [5:29] simple prompt to ask the model to [5:31] understand our training script and [5:33] optimize the model through a few [5:34] training attempts. Because of cursor [5:37] rule settings, the model will [5:38] automatically run tasks on the Domino [5:41] platform using the MCP server tools [5:43] available to it. [5:46] We're using Claude 3.7 thinking here and [5:49] it first reads the training script, the [5:51] model architecture, and the Domino [5:53] project settings. [5:55] After verifying git is synchronized [5:57] between our IDE and Domino, it runs a [6:00] Domino job with a first training [6:02] attempt. 10 epochs, 32 batch size, a [6:06] small model, and 0001 learning rate. [6:11] After checking that the job is [6:13] completed, it retrieves the results. And [6:15] here you can see that it sees we've [6:16] achieved a decent accuracy of 96%. [6:20] It also opens a web browser showing the [6:22] specific experiment run that was logged [6:25] on Domino from this training attempt. [6:33] While we're on the Domino UI, we can see [6:35] here that cursor ran the training like [6:36] any other domino job, meaning it's [6:39] tracked and fully reproducible by any [6:41] project collaborators. [6:45] Okay, so Cursor has decided it's going [6:47] to make another training attempt. This [6:49] time, slightly more epochs, a bigger [6:51] model, and reduced learning rate. [6:56] I've zoomed in to show you we're getting [6:57] a higher accuracy on this second [6:59] attempt. But what's cooler is that [7:01] Cursor has decided it's worried about [7:03] overfitting and is proposing to make a [7:05] code change to the model architecture to [7:08] increase the dropout rate slightly. [7:13] After checking the results of this third [7:15] run, cursor has decided to make one last [7:17] change to our model code, adding weight [7:19] decay to our atom optimizer. [7:24] Checking the results of this final [7:25] domino job run, cursor is providing a [7:28] nice summary about the training run [7:29] attempts it's done and has concluded the [7:31] final attempt is a good enough model [7:33] candidate at over 97%. [7:36] The key thing is everything we've done [7:38] here is fully tracked and reproducible [7:40] from the Domino platform. Our modeling [7:43] assistant is a safer and more [7:44] maintainable way to use AI agents in [7:46] your ML development workflow. [7:56] So this is this was a demo for Domino [7:58] modeling assistant basically providing [8:00] you a fully reproducible and full [8:02] lineage of the model that's running uh [8:04] within Domino and with with uh vibe [8:06] modeling. How how we have built this is [8:09] we uh we have technically have an IDE [8:12] which in this case is cursor uh a coding [8:14] assistant deployed on Domino MCP [8:17] servers. uh we have Domino APIs that are [8:19] automating compute, data access and job [8:21] orchestration for you and then Domino [8:24] system of record captures all the [8:25] actions, results and artifacts [8:27] forability purposes and giving you the [8:29] full lineage of that. This is uh sort of [8:31] like sort of like the architecture that [8:34] we have put together for the for Domino [8:36] modeling uh Y modeling code. [8:41] Some of the examples uh to think about [8:44] where we can where we have implemented [8:46] this is one data preparation. So you can [8:48] explore a data set uh generate [8:50] transformation script summarize apply [8:52] changes. Second one is model tuning um [8:55] basically you can iterate architectures [8:57] training configs track metrics in ML [9:00] flow and then debugging a rag app [9:03] retrieving logs inspect failures create [9:05] workspace. So these are some of these [9:06] are like the three main use cases that [9:09] we have really created the D modeling [9:11] assistant for now but it is it is just [9:12] the start for us and we are building it [9:15] uh for more we have we have things on [9:17] the road map that we are building on for [9:19] with the assistant. So with this I think [9:22] one of the cool uh one of the cool [9:24] things about Domino modeling assistant [9:26] is it's providing you flexibility and [9:28] power one to do the complete [9:29] orchestration uh which includes having a [9:32] IDE flexibility being ID agnostic you [9:35] can run it on any IDE and easier [9:38] deployment of your apps and APIs and API [9:40] based platform automation and next from [9:43] a from a concept of central system of [9:46] record you have central repository of [9:47] all the materials that are being [9:49] produced as part of the model training [9:51] one is giving you the entire [9:52] reproducibility, experiment tracking, [9:54] model registry, reusable assets and all [9:57] that. So those are the core advantages [9:59] that you get along with having control [10:01] and safe safety that you don't have to [10:03] sacrifice on uh when you think about W [10:05] modeling [10:08] Domino model assistant just to just [10:09] reiterate uh we were demonstrating that [10:11] using the Domino core platform [10:13] capabilities and this is what domino [10:15] platform is uh we are uh if you look at [10:18] the experiment tracking and [10:19] reproducibility concept orchestration [10:21] which leverages all of your [10:23] infrastructure data orchestration [10:24] capabilities underneath and entire model [10:26] life cycle orchestration and uh take an [10:29] idea to production very quickly and then [10:31] all of this uh is built with core [10:33] governance which helps you uh which [10:35] helps you operates within our secure [10:37] arbback platform. Uh as well uh on the [10:40] storage perspective definitely we have [10:42] intelligent data infrastructure [10:44] underneath supported by NetApp. So uh [10:46] FSX bundle on top that is something that [10:48] we support today. We have our data sets [10:51] capability that's that's on on NetApp. [10:53] So with with net app volumes uh you can [10:55] uh technically unlock the scalable and [10:58] data access with domino volumes that we [11:00] have created for this. [11:02] With this I think we are uh we are we [11:05] are definitely on I just wanted to [11:07] quickly call out some of the some of the [11:09] materials that you can tech take a look [11:11] at the LinkedIn blog post based on which [11:13] I have built this presentation. Uh we [11:15] also have a wide modeling blog that you [11:17] can take a look at. We have our Domino [11:20] MCP server available on uh GitHub uh [11:23] repo that's that's open and public and [11:25] definitely you know you can reach out to [11:27] us on uh partners domino.ai/artersnetapp [11:29] AI/Wartness NetApp. We are at four 415 [11:32] booth. So do visit us. Uh we can [11:34] demonstrate you the complete platform as [11:36] well. And yeah, with this uh this is my [11:38] time and I'm open for any questions that [11:40] you may have. [11:49] >> So tech are agnostic in that nature. Um [11:53] but with NetApp what we have done is we [11:56] have created a native experience for you [11:58] to be able to use those capability right [12:00] at Yeah. [12:14] >> Any other questions? [12:15] >> Any other questions? [12:20] >> No. Uh we are a separate company. is we [12:22] have we have partnered with NetApp. Uh [12:24] so we that's what it is. Yeah. [12:33] Any questions around modeling or coding [12:36] pieces of this? Happy to take those. [12:46] Actually [12:48] it is not but the purposes of both those [12:50] things are different. I think uh data [12:52] science and model building life cycle is [12:55] different from what we build in software [12:57] engineering right because it's static [12:58] code once you deploy it doesn't change [13:00] with data and models do so so being able [13:03] to track a lot of the work that you're [13:04] doing with models after you deploy it to [13:06] them in production the work starts [13:07] really after that the retraining and [13:09] being able to track all the history I [13:11] think it's really building it for the [13:13] model needs instead of a software [13:15] application needs that's where the [13:16] difference is Cool. [13:26] Uh so if there aren't any other [13:28] questions, do drop by on our booth. Uh [13:30] we'll be there to demonstrate you the [13:31] Domino and and Y modeling again. And [13:33] thank you. Thank you so much for the [13:34] time today. Uh really appreciate that. [13:36] Thank you.