AI Practices AI实践 2d ago Updated 2d ago 更新于 2天前 46

Developing NVIDIA Holoscan Applications with CLI, Skills, and AI Coding Agents 使用CLI、Skills和AI编码代理开发NVIDIA Holoscan应用程序

NVIDIA demonstrated an AI coding agent-assisted workflow for building real-time endoscopic tool segmentation applications using Holoscan and HoloHub The iterative development approach decomposes complex goals into smaller, verifiable engineering iterations guided by engineer-defined objectives and constraints An ablation study showed that combining Holoscan CLI, development skills, and documentation/examples produced the most efficient and highest-quality development workflow compared to using d 使用NVIDIA Holoscan CLI、HoloHub和AI编码代理(如Codex)协作开发实时内窥镜工具分割应用,实现从模型推理到可视化渲染的完整流水线 工程师通过迭代式目标定义引导AI代理开发,每轮迭代包含实现、测试、基准测试和优化,显著提升吞吐量并降低延迟 消融研究证明,结合CLI、skills和文档/示例的组合资源比单独使用文档或CLI能带来最高效、最高质量的开发工作流 开发流程采用渐进式分解策略,将复杂目标拆分为可验证的迭代步骤,确保设计选择能及时审查 Holoscan作为边缘AI应用平台配合HoloHub参考库,为医疗影像、机器人等实时AI应用提供标准化开发工具链

62
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • NVIDIA demonstrated an AI coding agent-assisted workflow for building real-time endoscopic tool segmentation applications using Holoscan and HoloHub
  • The iterative development approach decomposes complex goals into smaller, verifiable engineering iterations guided by engineer-defined objectives and constraints
  • An ablation study showed that combining Holoscan CLI, development skills, and documentation/examples produced the most efficient and highest-quality development workflow compared to using documentation or CLI alone
  • The agent-agnostic workflow was tested with Codex using GPT-5.6 sol max mode, completing the first iteration in approximately 40 minutes of agent processing time
  • The Holoscan CLI provides a shared execution interface that both the AI agent and human engineer can use, enabling inspection and repeatability of development operations

Why It Matters

This represents a significant step toward practical AI-assisted software development for specialized edge AI platforms, demonstrating that general-purpose coding agents can effectively navigate complex SDK ecosystems when provided with the right tools and documentation. For AI practitioners and edge computing developers, this workflow pattern could dramatically reduce the time and expertise required to build production-grade real-time AI applications in domains like medical imaging and robotics.

Technical Details

  • Platform: NVIDIA Holoscan for real-time edge AI applications, with HoloHub as the companion repository containing reference applications and components
  • Development Tools: Holoscan CLI (invoked via ./holohub wrapper), HoloHub development skills (holohub-app-lifecycle, holohub-debug-build-run), and progressive disclosure documentation via agents.md
  • Application Target: Real-time endoscopic tool segmentation reusing an existing MONAI model, with added HoloViz overlay for mask visualization, runtime telemetry, and repeatable benchmarking
  • Workflow Architecture: Engineer defines goals and constraints → AI coding agent inspects examples and documentation → Agent implements code and runs CLI operations → Engineer reviews outputs and sets next iteration goals
  • Benchmarking Approach: The agent inspected reference applications (monai_endoscopic_tool_seg, endoscopy_tool_tracking, surgical_scene_recon), generated scaffolding via CLI, implemented the application graph using existing Holoscan operators, and connected video replay, preprocessing, TensorRT inference, segmentation postprocessing, telemetry, and visualization

Industry Insight

  • The agent-agnostic nature of this workflow suggests that organizations can adopt AI coding assistants across diverse development teams without being locked into a single agent platform, enabling gradual integration of agentic tools into existing engineering pipelines
  • The ablation study findings reinforce the importance of providing AI agents with structured, tool-accessible documentation and executable interfaces rather than raw text documentation alone, which should inform how SDK vendors design their developer experience and agent support
  • The iterative decomposition strategy—breaking complex applications into reviewable, evidence-guided iterations—offers a practical template for managing AI-assisted development risk, ensuring human oversight at each stage while leveraging agent speed for implementation

TL;DR

  • 使用NVIDIA Holoscan CLI、HoloHub和AI编码代理(如Codex)协作开发实时内窥镜工具分割应用,实现从模型推理到可视化渲染的完整流水线
  • 工程师通过迭代式目标定义引导AI代理开发,每轮迭代包含实现、测试、基准测试和优化,显著提升吞吐量并降低延迟
  • 消融研究证明,结合CLI、skills和文档/示例的组合资源比单独使用文档或CLI能带来最高效、最高质量的开发工作流
  • 开发流程采用渐进式分解策略,将复杂目标拆分为可验证的迭代步骤,确保设计选择能及时审查
  • Holoscan作为边缘AI应用平台配合HoloHub参考库,为医疗影像、机器人等实时AI应用提供标准化开发工具链

为什么值得看

本文展示了AI编码代理如何与工程师协作完成复杂的边缘AI应用开发,为AI辅助软件工程提供了可复现的实践范式。对于关注AI开发工具链、边缘计算和医疗AI应用的从业者,本文提供了从理论到落地的完整技术参考。

技术解析

  • Holoscan CLI与HoloHub协作机制:通过./holohubwrapper调用Holoscan CLI作为共享执行接口,AI代理和工程师使用相同的CLI命令进行开发操作。HoloHub提供渐进式披露的文档模式(通过agents.md)和开发技能(如holohub-app-lifecycleholohub-debug-build-run),引导代理完成应用生命周期管理。
  • 迭代式AI辅助开发工作流:采用"工程师定义目标→AI代理实现→工程师审查→下一轮迭代"的循环模式。每轮迭代产生可审查的代码、输出和测试,为后续迭代提供依据。示例中使用Codex with GPT-5.6 sol max模式,首轮开发耗时约40分钟。
  • 端到端内窥镜工具分割应用架构:应用复用现有MONAI分割模型和Holoscan示例视频,连接视频回放、预处理、TensorRT推理、SDK分割后处理、遥测和HoloViz可视化模块。实现实时推理、掩码可视化、覆盖率/时间线显示和不确定性测量。
  • 消融研究设计:对比三种开发资源组合(仅文档、仅CLI、CLI+skills+文档/示例),验证组合资源在开发效率和代码质量上的优势,为AI辅助开发工具设计提供实证依据。

行业启示

  • AI编码代理的工程化落地路径:本文验证了"工程师定义约束+AI执行实现"的协作模式在复杂AI应用开发中的可行性,为AI辅助开发工具的设计提供了可复用的方法论框架。
  • CLI作为人机协作的共享接口价值:统一的CLI接口使工程师能够审查、复现和调试AI代理的操作,降低了AI辅助开发的可控性和可追溯性门槛,值得在更多AI开发工具中推广。
  • 边缘AI应用开发的标准化趋势:Holoscan平台通过CLI、skills和参考应用的组合,降低了实时AI应用(尤其是医疗影像等关键领域)的开发门槛,预示边缘AI开发工具链正朝着标准化、模块化方向演进。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Code Generation 代码生成 GPU GPU Healthcare AI 医疗AI Deployment 部署