Amazon Robotics Uses AI Agents to Cut CAD Review Loops to Two Hours

AI for Hardware Development With Amazon Robotics — Ai4

Ai4

Connecting multimodal AI directly to 3D CAD via Model Context Protocol catches 70% of manufacturing defects in two hours, bypassing multi-week supplier review bottlenecks before tooling begins.

AI-assisted summary of AI for Hardware Development With Amazon Robotics — original source. Check the original for full context.

Source published: 2026-08-15

Detailed notes

## Background Published on YouTube by the channel Ai4 on August 15, 2026, this presentation features Satvik Sabarad, an Advanced Manufacturing Engineer at Amazon Robotics. The talk details how multimodal AI agents, Claude Code CLI, and custom Model Context Protocol (MCP) servers bridge directly into 3D CAD software to fuse symbolic geometric data with visual context, automating manufacturability feedback and should-cost modeling.

## TL;DR Satvik Sabarad outlines an AI-driven framework deployed upstream at Amazon Robotics to compress multi-week supplier review cycles into automated analyses taking mere hours. The architecture pairs Claude Code CLI with SolidWorks via custom MCP servers, generating real-time VBScript macros and capturing multi-angle renders to evaluate geometric clearances, tooling constraints, and cycle times. Modular Markdown instruction files orchestrate deterministic calculations alongside multimodal visual reasoning, compiling comprehensive HTML reports with flagged DFM violations. In pilot studies on injection-molded components, the system achieved 70% coverage of vendor feedback, 85% true-positive accuracy, and should-cost estimates within ±25% of supplier quotes.

## Atmosphere & style The presentation carries a crisp, pragmatic, and highly technical tone grounded in the real-world complexities of industrial manufacturing rather than speculative AI hype. Sabarad speaks with the authoritative, steady confidence of an engineer managing deployed systems at massive scale, seamlessly linking low-level CAD mechanics with agentic orchestration. The accompanying terminal and CAD demo evokes a compelling sense of tangible productivity, illustrating a practical bridge between cutting-edge LLMs and traditional mechanical engineering workflows.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Physical AI focuses on embodied intelligence in robotics, whereas AI for hardware development leverages AI upstream to invent, design, optimize, and manufacture physical hardware. - [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Hardware development cycles typically span 3 to 4 years due to multi-week supplier review delays, contract negotiations, and iterative redesigns for tooling and manufacturability. - [03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) External contract manufacturers typically require 8 to 10 weeks to return actionable Design for Manufacturing (DFM) feedback, creating a major program bottleneck. - [05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) Subtle design flaws like improper bend radii, small punched hole diameters, and non-standard tooling drive up manufacturing failure rates, tooling wear, and production costs. - [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) Using Model Context Protocol (MCP) and Claude Code CLI enables AI agents to interface directly with desktop CAD software like SolidWorks rather than relying on fragile hardcoded APIs. - [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) The agent dynamically synthesizes VBScript macros inside the CAD environment to extract symbolic feature trees while simultaneously capturing multi-angle rendered screenshots for multimodal evaluation. - [12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) Modular Markdown instruction files act as domain-specific agents for metals, plastics, and PCBAs, while an orchestrator manages the boundary between deterministic code and LLM reasoning. - [13:02](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=782s) Multimodal LLMs evaluate components holistically, understanding functional intent, structural relationships, and human assembly access beyond rigid rule checks. - [15:10](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=910s) Multimodal visual reasoning enables should-cost estimation by predicting raw material needs and machining cycle times directly from 3D geometry. - [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) In a plastic injection molding pilot, the AI agent achieved 70% DFM coverage and 85% accuracy in just 2 hours compared to a multi-week supplier turnaround. - [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) AI-driven should-cost modeling landed within ±25% of finalized supplier quotes, with ongoing fine-tuning targeting regional and vendor-specific profiles.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Physical AI vs. AI for Hardware Development

Satvik Sabarad opens his presentation by introducing his role as an Advanced Manufacturing Engineer at Amazon Robotics, where he interfaces directly between internal design teams and an extensive network of external contract manufacturers and vendors. While Amazon Robotics develops autonomous systems strictly for internal fulfillment and warehouse logistics rather than consumer markets, the operational scale is staggering: the organization has recently deployed over one million robots across warehouse facilities in the United States, Europe, and Southeast Asia.

Setting the conceptual scope for the session, Satvik draws a clear distinction between physical AI and "AI for hardware development." While physical AI focuses on robotics and the embodiment of intelligence inside physical mechanical systems, his work tackles the reverse problem: leveraging artificial intelligence to invent, design, optimize, and manufacture the physical hardware itself. Every tangible physical product, whether a smartphone, a laptop, an office chair, or the carpet underfoot, demands an exhaustive, multi-year product development lifecycle. A phone held in someone's hand today likely began initial design phases three or four years earlier. Moving from concept design to consumer deployment involves prototype fabrication, extensive testing, contract negotiations with global suppliers, vendor sourcing, cost modeling, and recurring iterative redesigns to fit vendor manufacturing capabilities.

- **Deployed Scale:** Amazon Robotics has crossed the milestone of 1,000,000+ active robots deployed globally across US, European, and Southeast Asian fulfillment networks. - **Hardware Development Timelines:** Modern consumer electronics and robotics require 3 to 4 years from initial design kick-off to full-scale commercial manufacturing. - **Focus Areas:** Deploying multimodal LLMs upstream to compress hardware cycles across two primary disciplines: Design for X (DFX)—spanning manufacturability, scalability, and reliability—and cost-optimization case studies based on live internal robot programs.

---

[03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) The Vendor Feedback Bottleneck and DFX Reviews

When scaling a complex robotic platform composed of thousands of unique custom components, the traditional engineering design review process creates a massive logistical and operational bottleneck. Hardware design engineers package complete engineering releases—including native 3D CAD models and detailed 2D manufacturing drawings—and distribute them to external suppliers across the globe. Contract manufacturers must parse every drawing and feature tree by hand to evaluate whether parts can be fabricated reliably at scale and whether tooling investments can be justified.

```mermaid flowchart TD accTitle: Traditional vs AI Upstream Hardware Review Flow subgraph Traditional["Traditional Supplier Feedback Cycle (8-10 Weeks)"] A["CAD & 2D Drawings Package"] --> B["Global Contract Manufacturer Review"] B --> C["Tooling & DFM Iteration Notes"] C --> D["Engineering Redesign Loop"] end

subgraph AIUpstream["Upstream AI Agent Review (Minutes)"] E["Native 3D CAD File"] --> F["Custom MCP Server & CLI Agent"] F --> G["Instant DFM, Tooling & Assembly Rules"] G --> H["Optimized Design Before Supplier Hand-Off"] end ```

The friction in this cycle stems from deep physical constraints: metal die castings, sheet metal bends, and injection-molded parts each demand custom hard tooling, strict draft angles, precise material specifications, and machine-specific clearances. Because vendor manufacturing capabilities differ wildly depending on machinery and geographical facilities, suppliers take anywhere from 8 to 10 weeks to return actionable Design for Manufacturing (DFM) feedback. If critical manufacturing, assembly, or maintainability flaws are caught only after a multi-month feedback delay, the entire product launch slips. By deploying multimodal AI directly upstream into the engineering design phase, Amazon Robotics aims to catch tooling clashes, inadequate bend radii, and assembly interferences in minutes rather than months.

---

[05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) Anatomy of Supplier DFM Feedback on Sheet Metal

To illustrate the exact nature of these reviews, Satvik shares a concrete example of an industrial sheet metal bracket undergoing standard vendor scrutiny. When an experienced manufacturing engineer examines the part, they flag multiple subtle design flaws that dictate production yields and long-term durability:

- **Corner Relief and Bend Radii:** The inner corner radius of a bend was modeled smaller than the recommended tool radius, creating excessive localized stress that causes the sheet metal to crack during stamping or after cycling in the field. - **Hole Diameter to Sheet Thickness:** A punched through-hole featured a diameter smaller than the sheet metal stock thickness, which causes high tooling wear, punch breakages, and micro-fractures around the hole perimeter. - **Tooling Consolidation and Cost Reduction:** By standardizing and consolidating distinct hole patterns into a single uniform diameter across the sheet, the supplier avoids purchasing custom punching dies or dedicated secondary machining setups. - **Design for Maintainability and Assembly (DFMA):** A recessed access hole forced factory technicians to blindly thread tools into cramped, inaccessible internal pockets during assembly. Relocating or expanding the hole eliminated manual assembly bottlenecks on the plant floor.

In large-scale robotics programs featuring thousands of unique parts, having human engineers manually review every single feature across hundreds of supplier packages creates an unsustainable cognitive burden, resulting in the multi-week delays typical of traditional hardware cycles.

---

[07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) The Technical Architecture: Claude Code, MCP, and Multimodal CAD

Compressing this multi-week cycle down to automated real-time evaluation relies on two major technological breakthroughs: the Model Context Protocol (MCP) introduced in late 2024, and rich multimodal visual analysis. Rather than relying on fragile, hardcoded REST API endpoints where engineers must explicitly map every incoming and outgoing parametric boundary, MCP allows an agent to interface directly with desktop CAD software like Dassault Systèmes SolidWorks. The AI agent inspects the running CAD session, dynamically determines what geometric context it requires, and actively writes its own code on the fly.

```mermaid flowchart LR accTitle: Multimodal CAD Ingestion Architecture subgraph CADBridge["CAD Extraction Layer"] A["SolidWorks CAD Session"] --> B["Custom MCP Server"] B --> C["On-the-Fly Macro / VBScript Generator"] C --> D["Symbolic Data: Bounding Box, Features & Material"] A --> E["Multi-Angle Rendered Screenshots (Top, Side, ISO)"] end

subgraph LLMAnalysis["Multimodal Reasoning Layer"] D --> F["Claude Code CLI Engine"] E --> F G["Engineering DFM Rule Checker"] --> F F --> H["Markdown DFM Report & Actionable Fixes"] end ```

When connecting an LLM agent via Claude Code CLI to SolidWorks over MCP, the agent can query whether a model is open. If a requested tool does not exist within the MCP server to parse a complex feature tree, the agent dynamically synthesizes custom Visual Basic (VBScript) macros, runs them inside the SolidWorks API environment, and extracts part metadata, material specifications, feature hierarchies, and geometric bounding boxes. Simultaneously, the system commands the CAD engine to capture multi-angle viewport renders—top, side, front, and isometric projections.

Combining symbolic engineering data (exact feature trees, tolerances, and material properties) with visual spatial renderings overcomes the brittleness of traditional deterministic rule checkers. The multimodal model evaluates spatial clearances and human assembly accessibility just like "a seasoned manufacturing engineer who is looking at the part in real time." At the end of each Claude Code CLI inspection loop, the session outputs a structured Markdown log detailing all flagged DFM violations, tooling cost drivers, and proposed geometry modifications.

[12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) Orchestrating Agents and Rules via Markdown Instructions

When scaling the analysis to new components, the entire workflow can be kicked off simply by invoking an existing Markdown instruction file that orchestrates the necessary steps. This modular framework easily extends across diverse hardware domains, including metals, injection-molded plastics, electronics, and PCBAs (Printed Circuit Board Assemblies). By encoding domain-specific logic into separate Markdown files, each file effectively acts as a specialized agent. An overarching orchestrator coordinates these agents and delegates subtasks as needed.

```mermaid flowchart TD accTitle: Markdown-Driven Agent Orchestration Architecture accDescription: Diagram showing an orchestrator routing CAD analysis tasks between deterministic scripts and LLM multimodal reasoning.

subgraph Inputs["Inputs & Orchestration"] CAD["3D CAD / PCBA Artifacts"] MD["Markdown Agent Instructions"] ORCH["Claude Code Orchestrator"] end

subgraph Execution["Execution Pathways"] DET["Deterministic Code / CAD Macros (Feature tree, geometry, rules)"] LLM["Multimodal LLM Reasoning (Visual context, component description)"] end

subgraph Outputs["Final Deliverables"] REP["HTML DFM & Cost Report"] end

CAD --> ORCH MD --> ORCH ORCH -->|Rule Checks & Measurements| DET ORCH -->|Visual & Semantic Analysis| LLM DET --> REP LLM --> REP ```

Crucially, this structure governs the switching mechanism between deterministic rule-based code and LLM-driven reasoning. When an evaluation requires exact geometric math or known programmatic CAD macros, the system invokes deterministic scripts; it only falls back to calling the LLM when qualitative understanding, contextual synthesis, or visual reasoning is required.

- Domain-specific Markdown files serve as modular agents for metals, plastics, electronics, and PCBAs. - The orchestration layer handles the decision boundary between deterministic execution and LLM inference. - Eliminates unnecessary model calls by keeping hard-and-fast geometric calculations in deterministic code.

---

[13:02](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=782s) Live DFM Workflow and Holistic Component Understanding

Demonstrating the implementation via a video recording sped up roughly 4x to 5x, the workflow pairs a 3D CAD model on the left with Claude Code running in the terminal on the right. Once the user loads the Markdown agent files into the Claude Code session and prompts it to perform a Design for Manufacturability (DFM) analysis, the agent dynamically generates a structured task list. It systematically executes CAD macros to extract dimensions, captures rendered visual screenshots from multiple perspectives, synthesizes a comprehensive component description, evaluates the part against explicit DFM guidelines, and compiles the entire output into an interactive HTML report.

The qualitative component description produced in the report highlights a key capability unique to multimodal LLMs: rather than merely processing disconnected dimensions through rigid rule checks, the model understands the part holistically, grasping its intended design function and structural features. Engineers can interrogate the model with design-specific queries, review visual callouts of flagged defects, and inspect a summary table detailing passing and failing manufacturing parameters. These DFM rules can be adjusted dynamically on the fly to reflect specific vendor capabilities, manufacturing regions, or internal Amazon Robotics engineering standards.

- The agent builds an autonomous task checklist: macro execution, visual capture, holistic description, and guideline evaluation. - Outputs compile cleanly into an HTML report highlighting flagged defects and parameter pass/fail metrics. - Rule sets remain fully configurable on the fly according to local supplier capabilities or internal corporate requirements.

---

[15:10](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=910s) AI-Powered Should-Cost Modeling

Beyond manufacturability, multimodal AI provides a quantitative framework for should-cost estimation. When external suppliers submit quotes, engineering teams historically had to take those figures at face value without an easy way to verify fair market pricing. By deploying multimodal models directly on the geometry, the team estimates baseline manufacturing costs for individual parts and scales those projections up to program-level bill of materials (BOM) costs to build an entire robot at target production volumes.

The underlying calculation breaks down into raw material costs and manufacturing processing costs. Manufacturing cost in machining workflows hinges primarily on estimated cycle time multiplied by the shop's hourly machine rate. Because cycle time is tightly coupled to part geometry, feature depth, and operational complexity, multimodal AI visual reasoning plays an essential role in arriving at realistic production times.

---

[16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) Pilot Evaluation Metrics and Real-World Results

To ensure model predictions remain grounded, the team benchmarks AI performance against real engineering ground truth using strict evaluation metrics for both DFM feedback and should-cost modeling:

In a pilot case study focused on complex plastic injection molding components, an external supplier returned 90 distinct pieces of DFM feedback after their typical multi-week review cycle. The AI agent identified approximately 70% of those supplier-flagged issues in just 2 hours of automated analysis. Of the issues identified by the AI, 85% were verified as accurate, true-positive DFM risks.

For should-cost modeling, the AI's preliminary price predictions landed within a ±25% delta of the supplier quotes. Active engineering efforts are focused on bridging this remaining cost and accuracy gap by fine-tuning models on historical supplier feedback and regional manufacturing datasets to create supplier-specific evaluation profiles.

- **2 hours vs. 2 weeks:** Compressed supplier feedback turnarounds from multi-week waiting periods to roughly two hours. - **70% DFM coverage:** Caught nearly three-quarters of all feedback items identified by human tooling experts. - **85% DFM precision:** Verified true-positive accuracy across flagged injection-molding issues. - **±25% should-cost accuracy:** Preliminary cost estimates consistently matched supplier quotation bands.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Physical AI focuses on robotics and the embodiment of intelligence inside physical mechanical systems, his work tackles the reverse problem: leveraging artificial intelligence to invent, design, optimize, and manufacture the physical hardware itself. — Satvik Sabarad

> [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Deploying multimodal LLMs upstream to compress hardware cycles across two primary disciplines: Design for X (DFX)—spanning manufacturability, scalability, and reliability—and cost-optimization case studies based on live internal robot programs. — Satvik Sabarad

> [03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) Contract manufacturers must parse every drawing and feature tree by hand to evaluate whether parts can be fabricated reliably at scale and whether tooling investments can be justified. — Satvik Sabarad

> [03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) By deploying multimodal AI directly upstream into the engineering design phase, Amazon Robotics aims to catch tooling clashes, inadequate bend radii, and assembly interferences in minutes rather than months. — Satvik Sabarad

> [05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) The inner corner radius of a bend was modeled smaller than the recommended tool radius, creating excessive localized stress that causes the sheet metal to crack during stamping or after cycling in the field. — Satvik Sabarad

> [05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) A punched through-hole featured a diameter smaller than the sheet metal stock thickness, which causes high tooling wear, punch breakages, and micro-fractures around the hole perimeter. — Satvik Sabarad

> [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) The AI agent inspects the running CAD session, dynamically determines what geometric context it requires, and actively writes its own code on the fly. — Satvik Sabarad

> [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) The multimodal model evaluates spatial clearances and human assembly accessibility just like a seasoned manufacturing engineer who is looking at the part in real time. — Satvik Sabarad

> [12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) When an evaluation requires exact geometric math or known programmatic CAD macros, the system invokes deterministic scripts; it only falls back to calling the LLM when qualitative understanding, contextual synthesis, or visual reasoning is required. — Satvik Sabarad

> [13:02](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=782s) Rather than merely processing disconnected dimensions through rigid rule checks, the model understands the part holistically, grasping its intended design function and structural features. — Satvik Sabarad

> [15:10](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=910s) Manufacturing cost in machining workflows hinges primarily on estimated cycle time multiplied by the shop's hourly machine rate. — Satvik Sabarad

> [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) In a pilot case study focused on complex plastic injection molding components, an external supplier returned 90 distinct pieces of DFM feedback after their typical multi-week review cycle. The AI agent identified approximately 70% of those supplier-flagged issues in just 2 hours of automated analysis. — Satvik Sabarad

## Mentioned Amazon Robotics, Satvik Sabarad, Claude Code CLI, Model Context Protocol (MCP), Dassault Systèmes SolidWorks, SolidWorks API, VBScript, Design for Manufacturing (DFM), Design for X (DFX), Design for Maintainability and Assembly (DFMA), PCBAs (Printed Circuit Board Assemblies), Ai4

## Video description Satvik Sabarad - Advanced Manufacturing Engineer, Amazon Robotics

Hardware development across every industry — robotics, automotive, silicon, datacenters, and aerospace — remains constrained by multi-week supplier cycles required to obtain manufacturability feedback and cost estimates, and these cycles represent the single largest lever on program timelines. This talk examines what becomes possible when a multimodal AI agent is given direct access to a 3D CAD file — its feature tree, drawings, tolerances, and rendered views — and tasked with reasoning about how a part should be manufactured and what it should cost. I will present a working system built on Claude Code CLI, custom agents, and MCP servers that bridges directly into CAD, fusing symbolic geometric data with rendered visual context to reason about geometry, tolerances, and manufacturability in a manner comparable to a seasoned engineer. The talk will cover the underlying architecture, the evaluation framework used to benchmark model outputs against real engineering ground truth, and results from production programs in which multi-week feedback loops have been reduced to minutes. Attendees will leave with a clear technical view of how multimodal agents can operate on real engineering artifacts and a practical framework for applying these methods within their own hardware domain.

[Source](https://www.youtube.com/watch?v=RiNp_Vx44H8)

中文简报

Amazon Robotics利用AI Agent将CAD审查周期缩短至两小时通过模型上下文协议将多模态AI直接接入3D CAD，可在两小时内捕获70%的制造缺陷，并在开模前绕开长达数周的供应商审查瓶颈。 Satvik Sabarad概述了在Amazon Robotics上游部署的一套AI驱动框架，该框架将原本需要数周的供应商审查周期压缩为仅需数小时的自动化分析。该架构通过自定义MCP服务器将Claude Code CLI与SolidWorks相连接，生成实时VBScript宏并截取多角度渲染图，以评估几何间隙、模具约束和周期时间。模块化Markdown指令文件协调确定性计算与多模态视觉推理，生成标明DFM违规项的完整HTML报告。在针对注塑组件的试点研究中，该系统实现了供应商反馈70%的覆盖率、85%的真正例准确率，以及与供应商报价偏差在±25%以内的应有成本估算。 Physical AI侧重于机器人中的具身智能，而面向硬件开发的AI则利用上游AI来进行实体硬件的发明、设计、优化和制造。 (0:04) 由于数周的供应商审查延迟、合同谈判以及围绕模具和可制造性的反复重新设计，硬件开发周期通常长达3至4年。 (0:04) 外部合同制造商通常需要8至10周才能返回具有可操作性的面向制造的设计（DFM）反馈，形成了主要的流程瓶颈。 (3:26) 弯曲半径不当、冲孔直径过小和非标模具等细微设计缺陷，会导致制造缺陷率、模具磨损及生产成本上升。 (5:39) 利用模型上下文协议（MCP）与Claude Code CLI，AI Agent能够直接与SolidWorks等桌面CAD软件交互，而无需依赖脆弱的硬编码API。 (7:50)

Reading summary

Satvik Sabarad of Amazon Robotics presents an upstream framework for AI in hardware development, using multimodal large language models and the Model Context Protocol to automate mechanical engineering reviews. By inspecting CAD models directly, the system aims to compress multi-month supplier iteration cycles down to hours across large-scale robotics programs. Compressing the Multi-Week Supplier Bottleneck (3:26) Hardware development lifecycles typically span three to four years from initial concept to commercial deployment, constrained by deep physical manufacturing requirements such as hard tooling, draft angles, and machine-specific tolerances. In complex robotics fleets—such as Amazon Robotics' deployment of over one million fulfillment robots—design releases containing native 3D CAD files and 2D drawings are distributed to external contract manufacturers, who routinely require 8 to 10 weeks to return actionable Design for Manufacturing (DFM) feedback. Subtle physical design flaws caught late in this cycle create major schedule slips. In industrial sheet metal components, for example, inner bend radii modeled smaller than tooling limits cause localized cracking, punched holes narrower than stock thickness induce severe punch wear, and cramped access holes obstruct manual assembly on the factory floor. When human engineers must parse thousands of parts across hundreds of supplier packages, the cognitive load leads to recurring launch delays. Hybrid Multimodal CAD Architecture via MCP (7:50) To shift physical validation upstream, the engineering team connects Anthropic's Claude Code CLI directly to desktop Dassault Systèmes SolidWorks sessions using the Model Context Protocol (MCP). Rather than relying on rigid, pre-mapped REST APIs, the agent dynamically evaluates the running CAD model and writes custom Visual Basic (VBScript) macros on the fly. This enables on-demand extraction of symbolic engineering data, including feature trees, geometric bounding boxes, tolerances, and material properties. Simultaneously, the agent commands SolidWorks to render multi-angle visual captures, including top, side, front, and isometric projections. Combining symbolic feature data with spatial visual reasoning allows the model to assess qualitative clearances, structural context, and tool reachability in a manner comparable to an experienced human manufacturing engineer. Modular Markdown Orchestration and Cost Modeling (12:06) The analysis scales across distinct hardware domains—such as sheet metal, injection-molded plastics, and printed circuit board assemblies (PCBAs)—by encoding domain-specific inspection rules into modular Markdown instruction files. An overarching orchestrator interprets these instructions to manage the execution boundary: programmatic CAD macros handle deterministic geometric math, while the multimodal LLM is invoked only for visual reasoning, holistic component understanding, and contextual synthesis. Beyond DFM error flagging, this pipeline enables automated should-cost modeling. By visually evaluating part complexity and feature depths, the system estimates machining cycle times. Multiplying projected cycle times by regional hourly machine rates and adding raw material costs establishes baseline component and bill-of-materials (BOM) estimates, giving procurement teams an independent benchmark when reviewing supplier quotes. Empirical Pilot Benchmarks in Injection Molding (16:24) Amazon Robotics benchmarked the agent's performance against human supplier ground truth using three core metrics: DFM coverage (percentage of vendor issues detected), DFM accuracy (true-positive rate), and cost delta against finalized quotes. In a pilot study evaluating complex plastic injection molding components, an external vendor returned 90 DFM feedback items after a multi-week review. The automated agent evaluated the same components in two hours, capturing approximately 70% of the supplier-identified feedback issues with an 85% true-positive accuracy rate. For should-cost projections, preliminary AI-driven price estimates aligned within a ±25% delta of negotiated supplier pricing. To close the remaining coverage and cost accuracy gaps, Amazon Robotics is focused on fine-tuning models on historical supplier feedback records and regional manufacturing data to build vendor-specific evaluation profiles.

精炼阅读摘要

亚马逊机器人（Amazon Robotics）高级制造工程师 Satvik Sabarad 指出，硬件研发传统上面临长达数年的周期与数周的外部供应商制造可行性评审瓶颈；通过在设计上游引入多模态大语言模型与模型上下文协议（MCP），工程团队能够实现制造可行性（DFM）与应计成本（Should-Cost）的自动化实时评估。传统硬件研发瓶颈与制造评审摩擦 (3:26) Sabarad 强调，现阶段消费电子与工业机器人从概念立项到量产通常需要三至四年的研发周期。亚马逊机器人虽然仅服务于内部仓储物流，但其全球部署规模已突破一百万台。在由数千个定制零件组成的复杂机器人平台中，工程团队完成 3D CAD 模型与 2D 制造工程图纸后，需打包分发给全球代工厂商进行面向制造的设计（DFM）以及装配与可维护性设计（DFMA）审查。由于钣金折弯、压铸与注塑模具涉及严格的拔模角、折弯半径、开孔与板厚比以及定制工装夹具，各供应商的机器配置与制造公差差异巨大，通常需要八至十周才能返回评审反馈。以工业钣金支架为例，内角折弯半径小于刀具极限会导致冲压开裂，孔径小于料厚会引发冲头磨损，而复杂的盲孔装配设计则会严重阻碍产线装配。人工审查庞大的图纸包会造成严重的认知负荷与项目延期，迫使团队寻求在设计源头完成自动化缺陷拦截。结合 MCP 与多模态的 CAD 提取架构 (7:50) 为压缩数周的等待时间，该团队利用 2024 年末推出的模型上下文协议（MCP），将 Claude Code CLI 与达索 SolidWorks 等桌面 CAD 软件直接打通。与依赖硬编码 REST API 的传统方式不同，MCP 允许 AI 智能体直接探查运行中的 CAD 会话，并在现有接口工具缺失时，自主动态编写并执行 Visual Basic（VBScript）宏，从而实时提取零件特征树层级、材料属性、公差和几何包围盒等符号数据。在此基础上，系统通过调用 CAD 引擎自动捕获顶视、侧视、正视和轴测等多视角渲染截图。Sabarad 阐述道，传统确定性规则检查工具往往表现脆弱，而多模态模型能将精确的特征树参数与空间几何渲染图像相结合，像资深制造工程师一样直观评估零部件在三维空间中的间隙与装配可达性，最终输出包含违规标记与几何修改建议的结构化 Markdown 报告。模块化提示编排与确定性逻辑解耦 (12:06) 在多零件扩展层面，该架构利用结构化 Markdown 说明文件作为领域特定智能体，分别封装钣金、注塑、机加工以及印刷电路板组件（PCBA）等不同工艺的工程规则，并由中央编排器负责流程调度。这种设计的核心在于明确划分确定性代码与大模型推理的边界：对于精确的几何数学计算、特征尺寸测量与固定 CAD 宏，系统直接调用确定性代码执行；仅在涉及定性理解、空间干涉与视觉推理时，才调用多模态模型。在实机演示中，工程师将 Markdown 智能体文件导入 Claude Code 终端后，智能体会自动生成涵盖宏执行、图像捕获、零件综合描述与规则核验的执行清单。多模态模型展现出的独特优势在于能够全局理解零部件的结构功能与装配意图，而非单纯对比孤立的数字公差。此外，所有的 DFM 规则集均支持动态配置，工程师可根据特定供应商的工艺限制或特定区域的制造标准即时调整判断阈值。应计成本测算与复杂注塑件实测验证 (16:24) 除了制造缺陷审查，该系统还被用于建立量化应计成本（Should-Cost）模型，以解决过去外部供应商报价缺乏核验基准的问题。零件制造成本主要由原材料成本以及加工时间乘以机床小时费率构成，其中加工周期高度依赖于几何复杂度与特征深度，多模态 AI 的几何推理能力能够辅助推导出更贴近实际的加工耗时，进而外推至整机物料清单（BOM）的总成本。在针对复杂塑料注塑件的试点基准测试中，供应商耗时数周给出了 90 项具体 DFM 反馈，而 AI 智能体在两小时的自动化分析中成功捕获了其中约 70% 的问题（覆盖率），且被标记的问题中有 85% 经工程验证为真正的制造风险（准确率）。在成本预测方面，初步估算结果与供应商最终报价的偏差维持在正负 25% 以内，证实了上游快速评估的可行性。目前该技术仍处于试点演进阶段，正负 25% 的成本估算偏差表明模型与实际商业报价之间仍存差距。Sabarad 透露，团队当前的研发重点在于引入历史供应商反馈数据与区域制造数据集对模型进行针对性微调，以建立特定供应商与特定工厂的定制化评估档案，进一步提升复杂几何缺陷捕获的覆盖深度与精度。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2026 年 8 月 15 日发布，主讲人为亚马逊机器人（Amazon Robotics）先进制造工程师 Satvik Sabarad。视频简介指出，演讲重点探讨了如何利用 Claude Code CLI、自定义 Agent 以及 MCP 协议直接打通 3D CAD 环境，通过多模态 AI 融合几何特征与视觉渲染进行可制造性分析与成本估算，将原本长达数周的供应商评审周期缩减至数分钟。 ## TL;DR 针对传统硬件研发中跨地区代工厂反馈缓慢、DFM 与成本评估耗时长达数周的痛点，亚马逊机器人团队推出了一套结合 MCP 协议与多模态大模型的硬件自动化分析架构。该方案通过终端工具 Claude Code CLI 实时联动 SolidWorks CAD，动态生成执行脚本并抓取多视角渲染图像，借助 Markdown 规范将确定性工程规则与大模型视觉语义理解深度解耦并混合编排。系统不仅能自动输出高亮隐患特征的交互式 DFM 报告，还能基于生产节拍科学推导零件的应有成本（Should-Cost）。在注塑件生产试点中，团队成功将原本 2 周的供应商评审压缩至 2 小时，并达成了 70% 的缺陷检出覆盖率与 85% 的判定准确率。 ## Atmosphere & style 整场分享散发着极为务实硬核的工程底色与前沿极客气息，演讲者以资深制造工程师的严谨视角剖析供应链痛点，摒弃了空洞泛化的 AI 概念宣讲。伴随终端 Claude Code 快速交互与 CAD 实操演示的高效推进，现场呈现出一种软件敏捷迭代与重资产实体制造深度交融的确定感与工业冲击力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Satvik Sabarad 明确区分了物理 AI（Physical AI，具身智能）与面向硬件开发的 AI（AI for Hardware Development，用于加速工程研发全流程），亚马逊仓储机器人全球部署量已突破 100 万台。 - [01:09](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=69s) 传统硬件研发受制于跨地域代工厂漫长的供应商评审周期，开模、拔模角及工艺反馈往往需要等待 8 到 10 周，亟需在前序设计阶段引入 AI 进行 DFX 拦截。 - [05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) 一套大型机器人包含数千个非标零件，人工逐页审查钣金或铸件的折弯内角、孔径厚度比及装配通道等 DFM 细节，在多项目并行时面临严重的效率与规模瓶颈。 - [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) 系统基于模型上下文协议（MCP）在 AI 智能体与 SolidWorks 间构建双向桥梁，摆脱传统 API 硬编码限制，赋予 AI 现场编写并执行代码的动态灵活性。 - [08:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=530s) 多模态视觉推理融合了 3D 几何特征参数与 CAD 视窗多角度渲染截图，使大模型能够像资深制造工程师一样同时理解抽象参数与整体构型直观可行性。 - [10:49](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=649s) 工程师利用 Claude Code CLI 与 CAD 交互，当预设工具不足时，AI 会自主编写并注入 VBS 宏脚本抓取特征，并在会话后将缺陷整理为标准化 Markdown 报告。 - [12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) 系统采用 Markdown 文件定义并调度多 Agent 工作流，实现了几何计算宏等确定性代码与多模态 LLM 模糊泛化推理之间的动态分流编排。 - [13:02](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=782s) 通过全自动流水线，Claude Code 能在几分钟内解析 CAD 零件并编译输出直观的 HTML 交互式 DFM 评估报告，高亮显示潜在制造隐患。 - [14:09](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=849s) 多模态大模型赋予系统对零件在机械结构中作用的全局语义理解，且工程师可根据不同代工厂或地域的制造能力在 Markdown 中动态调整公差规则阈值。 - [15:10](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=910s) 应有成本（Should-Cost）模型通过多模态视觉感知零件拓扑复杂度，精准推算加工周期（Cycle Time）与工时费率，为团队提供独立于供应商报价的客观核算基准。 - [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) 团队建立严苛的闭环量化评测体系，以 DFM 覆盖率、DFM 准确率以及成本报价偏差（Cost Delta）三项核心指标对齐真实供应商人工评审真值。 - [17:23](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=1043s) 在塑料注塑件实测中，AI 将 2 周的供应商反馈周期压缩至 2 小时，检出 70% 的制造缺陷并达成 85% 准确率，成本估算偏差保持在 ±25% 以内。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) 个人背景与硬件开发 AI 的定位 Satvik Sabarad 目前在亚马逊机器人（Amazon Robotics）担任先进制造工程师（Advanced Manufacturing Engineer），其核心工作是协同内部设计团队，并与分布在全球各地的代工厂（Contract Manufacturers）及供应商紧密合作，共同推进机器人解决方案的规模化量产。与面向大众消费者的产品不同，亚马逊机器人主要专注于内部仓储物流场景的自动化运作。目前，亚马逊机器人在美国、欧洲以及东南亚等全球物流枢纽中规模化部署的机器人总数已经正式突破了 100 万台大关。在展开具体技术方案前，Satvik 首先厘清了两个容易混淆的概念：当前备受瞩目的“物理 AI”（Physical AI，即将人工智能具身化嵌入到机器人等物理实体系统中执行物理任务），与本次演讲的核心主题——“面向硬件开发的 AI”（AI for Hardware Development，即利用 AI 工具与智能体来设计、验证和加速物理硬件本身的工程研发全流程）。 - 制造工程师与内部设计团队、全球合同制造伙伴协同工作以实现规模化量产。 - 亚马逊仓储机器人目前在全球部署量已超过 100 万台。 - 核心区分：具身智能物理 AI（Physical AI）与研发提效 AI（AI for Hardware Development）。 --- ### [01:09](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=69s) 传统硬件研发周期的瓶颈与 DFX 优化日常生活中所接触到的每一件物理产品——从手机、笔记本电脑，到坐着的椅子、脚下的地毯——都必须历经严苛漫长的产品开发流程。例如用户手中正在使用的一部普通智能手机，其初始设计往往在三四年前就已经启动，历经数年周折才最终交付到消费者手中。硬件研发之所以极其耗时，是因为团队不仅要完成原型设计与测试，还必须跨越地理界限与各地的代工厂建立合同协议，开展物料采购寻源（Sourcing），并根据制造厂商的具体工艺极限进行反复的成本优化与设计迭代。随着大语言模型（LLM）与人工智能技术的日益成熟，工程团队得以将 AI 深度融入硬件开发流程中，从而大幅压缩研发周期。本次技术分享重点聚焦于两大方向：首先是广义的“面向 X 的设计”（DFX），涵盖面向制造的设计（DFM）、面向可扩展性的设计（DFS）以及面向可靠性的设计（DFR）。这些来自制造端的跨维度反馈通常极其分散，设计团队必须在产品定型量产前将其全部消化并融入图纸中。演讲中展示了来自亚马逊机器人内部真实量产项目的脱敏案例，直观呈现了 AI 介入后研发周期的实质性缩减。 --- ### [03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) 供应商评审与反馈闭环的漫长拉锯在传统的设计评审流程中，设计团队完成整体设计后会打包生成一套包含 3D CAD 模型与 2D 工程图纸的“设计包”（Design Package），并将其分发给各地的制造供应商。面对台下众多的硬件工程师同行，Satvik 详细拆解了这一协作过程：供应商工程人员必须逐一审查图纸中的每个零件细节，评估其可制造性以及是否具备支撑数百万台级别量产的能力。设计工程师必须完整吸收供应商提出的所有修改意见，而这种跨组织的往复沟通极其耗时。以金属铸造件和压铸（Die Casting）工艺为例，开发涉及极为复杂的开模与模具制造（Tooling）环节：设计人员必须根据供应商现场的拔模角（Draft Angle）要求、材料选型特性以及特定机床的加工能力不断修改图纸。由于不同供应商、不同工厂乃至不同地理区域的制造能力千差万别，获取一轮完整的可制造性反馈通常需要等待 8 到 10 周之久。如果能在产品设计周期的最上游（Upstream）部署 AI 智能体，直接在前期拦截并修正这些潜在制造缺陷，将为硬件研发与上线部署节省难以估量的时间。 --- ### [05:39](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=339s) 钣金件 DFM 实例：人工评审的复杂性与规模瓶颈以一个典型的金属钣金零件（Sheet Metal Part）为例，供应商在进行 DFM 审查时通常会提出多维度的严苛工程反馈： ```mermaid flowchart TD accTitle: 钣金件 DFM 多维度制造反馈 accDescription: 展示传统制造供应商对钣金零件提出的制造工艺、成本优化和装配维护性反馈。 subgraph DFM["制造工艺维度 (DFM)"] A["拐角弯曲内角"] --> A1["半径过小需加大 避免长期应力开裂"] B["通孔几何尺寸"] --> B1["孔径与板厚不匹配 需调整孔径比例"] end subgraph Cost["成本与工装优化"] C["特征孔位整合"] --> C1["合并不同孔径规格 减少冲压模具与专用夹具投入"] end subgraph DFA["装配与可维护性 (DFA/DFM)"] D["内部深处装配孔"] --> D1["作业员难以触及 需重新规划装配空间通道"] end ``` 在传统作业模式下，这类问题完全依赖资深工程师逐页审查 2D 图纸与 3D 几何模型。对于单一简单机构而言尚可应对，但一套大型仓储机器人系统通常包含数千个独一无二的非标零部件（Unique Parts）。当多个研发项目并行推进、产量呈指数级放大时，合同制造商端的人工审查工作负荷会变得极其庞大，导致反馈周期动辄拖延数周甚至数月。 - **弯曲半径缺陷**：若折弯拐角处的内圆角半径小于材料允许极限，零件在长期受力服役中极易产生疲劳裂纹。 - **孔径厚度比失调**：冲孔直径与板材厚度比例不合规范，需按冲压工艺准则放大孔径。 - **模具成本削减**：通过统一步骤和孔型规格，消除非标加工需求，直接削减高昂的模具（Tooling）开发与专用设备采购支出。 - **装配可达性限制**：针对组装工人难以触碰的深处安装孔位提出面向装配的设计（DFA）改进建议。 --- ### [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) 核心突破：MCP 协议桥接与多模态视觉推理为了打破数周的等待周期，系统依赖于两大底层技术突破：首先是 2024 年底迅速普及的模型上下文协议（MCP, Model Context Protocol）。通过构建 MCP 服务器，团队在 AI 智能体与原生 CAD 软件 SolidWorks 之间架起了一座双向通信的动态桥梁，使得 AI 能够直接读取零件的材料属性、特征树（Features）以及三维包围盒（Bounding Box）。针对业内常见的“MCP 与传统 API 区别”的讨论，Satvik 强调，传统 API 必须预先硬编码定义好输入输出端点和固定参数；而 MCP 则赋予了 AI 智能体自主理解用户意图、规划任务并“现场即时编写并执行代码（Build the code right on the fly）”的动态灵活性。其次是多模态视觉分析能力的融合。单纯获取几何特征参数无法完全反映零件的真实制造样貌，因此系统通过自动化流程抓取 CAD 视窗中零件的顶视图、侧视图等全方位渲染截图。将三维特征参数、材料属性与多视角视觉截图同时喂给多模态大模型，并结合工程规则校验器（Rule Checker）进行联合判定。AI 由此得以像一位坐在屏幕前、实时审视实物模型的“资深制造工程师”，既能理解抽象的几何参数，又能依靠直观视觉洞察整体构型与工艺可行性。 ```mermaid flowchart LR accTitle: 基于 MCP 与多模态的多维度 CAD 推理流程 accDescription: 展示通过 MCP 从 CAD 获取特征数据与视觉截图并输入 LLM 进行 DFM 推理的架构。 subgraph CAD["SolidWorks CAD 环境"] M["几何特征 / 材料属性 / 包围盒"] V["顶视/侧视等多角度渲染截图"] end subgraph Agent["MCP 桥接与数据整合"] MCP["MCP 协议服务 (动态生成宏/脚本)"] M --> MCP end subgraph Reasoning["多模态推理核心"] LLM["多模态 LLM 智能体 (模拟资深制造工程师)"] RC["工程规则校验器 (Rule Checker)"] MCP --> LLM V --> LLM RC --> LLM end LLM --> Out["分钟级 DFM 评估与设计修改建议"] ``` --- ### [10:49](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=649s) 智能体实战交互：动态生成脚本与自动化流程沉淀在底层执行层面，工程师通过终端工具（如 Claude Code CLI）与 CAD 平台开展交互。当工程师接入 MCP 并询问 AI 是否能识别当前 SolidWorks 中打开的模型时，AI 会实时确认模型状态。若工程师要求提取深层特征数据而预设 MCP 工具库尚不支持该特定接口，AI 会展现出极强的自主性——它不会直接报错退出，而是现场即时编写专用的 VBS 脚本与宏命令（Macros）注入 CAD 环境执行，动态抓取所需特征。紧接着，智能体自动调用截图指令采集多视角渲染画面，将提取到的几何参数与图像数据送入 DFM 规则检查器中进行全面比对。在 Claude Code 交互会话结束时，工程师只需发出一句指令：“你能根据我们刚才讨论的内容生成一份 Markdown 报告吗？”，AI 便会迅速将所有发现的缺陷、参数对比及修改建议沉淀为标准化、结构化的工程 Markdown 文档，从而将原本长达数周的人工供应商走查闭环，彻底转化为数分钟内即可复现的标准化自动化流程。 ### [12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) 基于 Markdown 与多 Agent 的确定性规则与大模型混合编排在对不同零件进行可制造性分析（DFM）时，整个工作流的核心是通过结构化的 Markdown 文件来充当 Agent 的定义与调度指令。当需要分析一个全新零件时，系统只需调用现有的 Markdown 文件，它便会基于预先定义好的工程逻辑调度所有环节。这种模式具有极高的扩展性，无论是金属件、塑料注塑件，还是电子元器件与 PCBA（印制电路板组件），都可以遵循相同的范式：将各自专属的工程规则和上下文沉淀在对应的 Markdown 文件中，并在顶层构建一个编排器（Orchestrator），按需唤起不同的专业子 Agent。这种架构的一大关键优势在于实现了“确定性代码”与“大语言模型（LLM）”之间的动态分流。在传统的规则校验和几何计算中，确定性的代码逻辑（如脚本或 CAD 宏）更加精准且高效；而在需要对整体上下文进行泛化推理或视觉理解时，才触发 LLM 调用。Markdown 文件巧妙地定义了这种在确定性规则方法与 LLM 模糊推理之间的路由机制，确保每一次计算和判定都落在最合适的工具链上。 ```mermaid flowchart TD accTitle: 硬件多 Agent 混合编排架构 subgraph Orchestration["编排层与零件输入"] A["CAD 零件文件输入\n(金属/塑料/PCBA)"] --> B["顶层编排器\n(Orchestrator)"] B --> C["Markdown Agent 规范\n(规则与提示词配置)"] end subgraph Execution["混合执行层"] C --> D{"判定调用类型"} D -- "几何计算 / 宏提取" --> E["确定性代码\n(Deterministic Rules / Macros)"] D -- "视觉 / 语义理解" --> F["多模态 LLM\n(Claude Code CLI)"] end subgraph Output["交付层"] E --> G["综合 DFM 评估报告\n(HTML 报表 / 成本模型)"] F --> G end ``` ### [13:02](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=782s) Claude Code CLI 驱动的 DFM 自动化演示与分析流水线 Satvik 展示了一段以 4 到 5 倍速快进的实际运行视频。屏幕一侧是待测零件的 CAD 模型，另一侧则是正在终端中全速运行的 Claude Code CLI。当工程师将设定好的 Markdown 指令加载到会话中，并向其发出针对该零件执行 DFM 分析的指令后，系统便立刻自主拆解出详细的任务列表（Task List）。流水线的运作一气呵成：Claude Code 会逐一执行提取宏定义、自动化截取 CAD 零件的多角度渲染图、撰写组件描述，并对照预先设定的 DFM 设计准则逐项进行严密评估。在极短的运行时间后，系统最终编译输出一份直观详尽的 HTML 格式 DFM 评估报告，将零件上存在制造隐患与违规特征的区域逐一高亮标出，极大解放了人工核对图纸的繁琐负担。 - **任务编排步骤**：自动拆解 Task List $\rightarrow$ 提取 CAD 几何宏参数 $\rightarrow$ 截取零件渲染视点 $\rightarrow$ 生成组件语义描述 $\rightarrow$ 对齐 DFM 准则评估 $\rightarrow$ 输出最终 HTML 交互式报告。 - **运行环境**：基于终端运行的 Claude Code CLI，直接连接 CAD 会话，通过多模态交互完成端到端自动化。 ### [14:09](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=849s) 软硬件基石：全局语义理解与动态规则配置在生成的 DFM 报告中，最引人瞩目的是对零件的“组件整体描述”（Component Description）。这是纯基于几何规则的传统几何引擎无法做到的——得益于多模态 LLM 的引入，系统不再把零件当成冷冰冰的面和边的组合，而是能够像人类资深工程师一样，对零件在机械系统中的结构作用和特征进行“全貌式理解”（Holistic Understanding）。这为后续更深度的工程问答和设计优化奠定了坚实的硬件认知基础。此外，报告末尾提供了清晰的汇总表格，全面列出各项制造公差与工艺参数的“通过（Pass）”与“失败（Fail）”状态。更具工程实用价值的是，这些校验规则完全不是写死的逻辑，工程师可以根据具体合作的代工厂加工能力、地理区域供应链差异，或是 Amazon Robotics 内部的设计规范，随时在 Markdown 中实时动态修改规则阈值，并在极短时间内完成重新评估。 - **多模态核心价值**：突破传统 CAD 规则检查的局限，具备对零件功能与装配关系的语义理解。 - **灵活的定制性**：支持针对特定供应商制程能力、制造地域、内部专属标准进行敏捷调整。 ### [15:10](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=910s) 基于制造节拍的多模态应有成本模型（Should-Cost）除了可制造性反馈，硬件研发团队面临的另一大痛点是“应有成本”（Should-Cost）的精准建模。在传统流程中，供应商给出的报价往往被默认为合理基准，但研发团队亟需独立的客观验证工具，以判断供应商报价是否存在虚高。Amazon Robotics 利用多模态 AI 深入解析零件的几何复杂度与工序特征，推导出单件加工成本，并最终向上扩展至整机在特定量产规模下的整机项目级成本估算。在底层工程计算上，零件成本主要由原材料成本与加工制造费用构成。以机加工（Machining）为例，加工费用直接取决于加工周期（Cycle Time）乘以机加车间的工时费率（Shop Hourly Rate）。估算加工周期的难点在于理解刀具路径、装夹次数以及复杂拓扑特征，这正是多模态 AI 结合三维模型与视觉感知大显身手的领域。通过让 AI 真正“看懂”零件的拓扑形态，系统能够更科学地倒推生产节拍与真实加工成本。 ### [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) 闭环评测体系：DFM 覆盖率与预测精度矩阵 Satvik 强调，团队绝不会“盲目相信 AI 给出的输出结果”（"We don't blindly believe in what AI is saying."），而是建立了一套极其严苛的量化评测体系。为了确保多模态系统能够对齐真实生产标准，Amazon Robotics 将模型评估细分为可制造性与成本两大核心维度，直接与供应商专家的实测反馈进行横向对比： | 评估维度 | 核心量化指标 | 具体计算逻辑与定义 | |:--- |:--- |:--- | | **可制造性 (DFM)** | **覆盖率 (DFM Coverage)** | 供应商人工反馈的全部制造缺陷中，被 AI 成功检出并标记的问题占比。 | | **可制造性 (DFM)** | **准确率 (DFM Accuracy)** | 在 AI 报告的所有潜在 DFM 缺陷中，经工程验证确实成立的真实问题占比。 | | **应有成本 (Should-Cost)** | **报价偏差 (Cost Delta)** | AI 自动化预测的零件成本与供应商实际商业报价之间的相对百分比误差。 | ### [17:23](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=1043s) 生产试点实测：塑料注塑件数周周期压缩至两小时在实际生产项目的试点验证中，团队挑选了一批极具代表性的复杂塑料注塑成型（Injection Molding）零件。在与真实供应商的闭环测试中，供应商的人工 DFM 分析团队共挑出了 90 项具体修改意见。与之对照，AI 工具在短短 2 小时内就捕获了其中约 70% 的制造缺陷；相比之下，供应商完成这套全面分析通常需要整整两周的往返沟通周期。在 AI 检出的缺陷项中，准确率达到了 85%（仅有 15% 为误报），目前团队正在通过引入往期历史供应商反馈数据对模型进行持续微调（Fine-tuning），以消除这 15% 的误判并贴合特定供应商的工艺风格。而在成本预测方面，AI 给出的应有成本估算值与供应商真实报价之间的误差保持在 ±25% 以内。随着微调与多模态数据链路的持续迭代，该成本差距正在被进一步压缩，使硬件工程团队能够在数分钟内获得媲美资深供应链专家的反馈见解。 - **反馈周期跃迁**：从传统供应商往返需要的 2 周大幅缩短至 AI 运行的 2 小时。 - **缺陷捕获表现**：在 90 项真实供应商反馈中，AI 实现了约 70% 的缺陷覆盖率（Coverage）。 - **缺陷判定精度**：已识别缺陷的准确度（Accuracy）达到 85%，当前正基于历史数据微调消除 15% 偏差。 - **成本估算偏差**：AI 预测与供应商真实报价差距稳定在 $\pm 25\%$ 区间，持续收窄中。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=4s) Physical AI 与面向硬件开发的 AI（AI for Hardware Development）是两个截然不同的概念，后者专注于利用 AI 工具与智能体来设计、验证和加速物理硬件本身的工程研发全流程。 — Satvik Sabarad > [03:26](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=206s) 由于不同供应商、不同工厂乃至不同地理区域的制造能力千差万别，获取一轮完整的可制造性反馈通常需要等待 8 到 10 周之久。 — Satvik Sabarad > [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) 传统 API 必须预先硬编码定义好输入输出端点和固定参数；而 MCP 则赋予了 AI 智能体自主理解用户意图、规划任务并现场即时编写并执行代码（Build the code right on the fly）的动态灵活性。 — Satvik Sabarad > [07:50](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=470s) AI 由此得以像一位坐在屏幕前、实时审视实物模型的资深制造工程师，既能理解抽象的几何参数，又能依靠直观视觉洞察整体构型与工艺可行性。 — Satvik Sabarad > [12:06](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=726s) Markdown 文件巧妙地定义了这种在确定性规则方法与 LLM 模糊推理之间的路由机制，确保每一次计算和判定都落在最合适的工具链上。 — Satvik Sabarad > [14:09](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=849s) 系统不再把零件当成冷冰冰的面和边的组合，而是能够像人类资深工程师一样，对零件在机械系统中的结构作用和特征进行全貌式理解（Holistic Understanding）。 — Satvik Sabarad > [16:24](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=984s) We don't blindly believe in what AI is saying. — Satvik Sabarad > [17:23](https://www.youtube.com/watch?v=RiNp_Vx44H8&t=1043s) 在与真实供应商的闭环测试中，供应商的人工 DFM 分析团队共挑出了 90 项具体修改意见。与之对照，AI 工具在短短 2 小时内就捕获了其中约 70% 的制造缺陷。 — Satvik Sabarad ## Mentioned Satvik Sabarad, Amazon Robotics, SolidWorks, Claude Code CLI, MCP (Model Context Protocol), DFM (Design for Manufacturing), DFS (Design for Scalability), DFR (Design for Reliability), DFA (Design for Assembly), Should-Cost ## Video description Satvik Sabarad - Advanced Manufacturing Engineer, Amazon Robotics Hardware development across every industry — robotics, automotive, silicon, datacenters, and aerospace — remains constrained by multi-week supplier cycles required to obtain manufacturability feedback and cost estimates, and these cycles represent the single largest lever on program timelines. This talk examines what becomes possible when a multimodal AI agent is given direct access to a 3D CAD file — its feature tree, drawings, tolerances, and rendered views — and tasked with reasoning about how a part should be manufactured and what it should cost. I will present a working system built on Claude Code CLI, custom agents, and MCP servers that bridges directly into CAD, fusing symbolic geometric data with rendered visual context to reason about geometry, tolerances, and manufacturability in a manner comparable to a seasoned engineer. The talk will cover the underlying architecture, the evaluation framework used to benchmark model outputs against real engineering ground truth, and results from production programs in which multi-week feedback loops have been reduced to minutes. Attendees will leave with a clear technical view of how multimodal agents can operate on real engineering artifacts and a practical framework for applying these methods within their own hardware domain. [Source](https://www.youtube.com/watch?v=RiNp_Vx44H8)

Source quotes

[0:04] Satvik Sabarad: Physical AI 与面向硬件开发的 AI（AI for Hardware Development）是两个截然不同的概念，后者专注于利用 AI 工具与智能体来设计、验证和加速物理硬件本身的工程研发全流程。 [3:26] Satvik Sabarad: 由于不同供应商、不同工厂乃至不同地理区域的制造能力千差万别，获取一轮完整的可制造性反馈通常需要等待 8 到 10 周之久。 [7:50] Satvik Sabarad: 传统 API 必须预先硬编码定义好输入输出端点和固定参数；而 MCP 则赋予了 AI 智能体自主理解用户意图、规划任务并现场即时编写并执行代码（Build the code right on the fly）的动态灵活性。 [7:50] Satvik Sabarad: AI 由此得以像一位坐在屏幕前、实时审视实物模型的资深制造工程师，既能理解抽象的几何参数，又能依靠直观视觉洞察整体构型与工艺可行性。 [12:06] Satvik Sabarad: Markdown 文件巧妙地定义了这种在确定性规则方法与 LLM 模糊推理之间的路由机制，确保每一次计算和判定都落在最合适的工具链上。 [14:09] Satvik Sabarad: 系统不再把零件当成冷冰冰的面和边的组合，而是能够像人类资深工程师一样，对零件在机械系统中的结构作用和特征进行全貌式理解（Holistic Understanding）。 [16:24] Satvik Sabarad: We don't blindly believe in what AI is saying. [17:23] Satvik Sabarad: 在与真实供应商的闭环测试中，供应商的人工 DFM 分析团队共挑出了 90 项具体修改意见。与之对照，AI 工具在短短 2 小时内就捕获了其中约 70% 的制造缺陷。

Source quotes

[0:04] Satvik Sabarad: Physical AI focuses on robotics and the embodiment of intelligence inside physical mechanical systems, his work tackles the reverse problem: leveraging artificial intelligence to invent, design, optimize, and manufacture the physical hardware itself. [0:04] Satvik Sabarad: Deploying multimodal LLMs upstream to compress hardware cycles across two primary disciplines: Design for X (DFX)—spanning manufacturability, scalability, and reliability—and cost-optimization case studies based on live internal robot programs. [3:26] Satvik Sabarad: Contract manufacturers must parse every drawing and feature tree by hand to evaluate whether parts can be fabricated reliably at scale and whether tooling investments can be justified. [3:26] Satvik Sabarad: By deploying multimodal AI directly upstream into the engineering design phase, Amazon Robotics aims to catch tooling clashes, inadequate bend radii, and assembly interferences in minutes rather than months. [5:39] Satvik Sabarad: The inner corner radius of a bend was modeled smaller than the recommended tool radius, creating excessive localized stress that causes the sheet metal to crack during stamping or after cycling in the field. [5:39] Satvik Sabarad: A punched through-hole featured a diameter smaller than the sheet metal stock thickness, which causes high tooling wear, punch breakages, and micro-fractures around the hole perimeter. [7:50] Satvik Sabarad: The AI agent inspects the running CAD session, dynamically determines what geometric context it requires, and actively writes its own code on the fly. [7:50] Satvik Sabarad: The multimodal model evaluates spatial clearances and human assembly accessibility just like a seasoned manufacturing engineer who is looking at the part in real time. [12:06] Satvik Sabarad: When an evaluation requires exact geometric math or known programmatic CAD macros, the system invokes deterministic scripts; it only falls back to calling the LLM when qualitative understanding, contextual synthesis, or visual reasoning is required. [13:02] Satvik Sabarad: Rather than merely processing disconnected dimensions through rigid rule checks, the model understands the part holistically, grasping its intended design function and structural features. [15:10] Satvik Sabarad: Manufacturing cost in machining workflows hinges primarily on estimated cycle time multiplied by the shop's hourly machine rate. [16:24] Satvik Sabarad: In a pilot case study focused on complex plastic injection molding components, an external supplier returned 90 distinct pieces of DFM feedback after their typical multi-week review cycle. The AI agent identified approximately 70% of those supplier-flagged issues in just 2 hours of automated analysis.

Original transcript

[0:04] Thank you. Thanks for making to this session. Um, I'm Satwick. I work at Amazon Robotics as a [0:10] manufacturing engineer. My work is with the design team within Amazon Robotics [0:15] and I work with a lot of contract manufacturers, vendors based out of uh different places [0:20] um in in ramping up the robotic solutions. uh primarily Amazon robotics is uh developing robots [0:27] for um internal warehouse applications. It's it's not for the consumers outside. [0:33] Um but we have just hit like a million robots um um deployed at scale at different locations within [0:41] the US in Europe in Southeast Asia. Um today I'll be speaking on AI for hardware development. [0:49] Um just wanted to set this stage a little bit. Uh this is a little different from physical AI. [0:54] Um physical AI is basically like you're embodying AI in a physical system like a robot or something. [1:00] But um using AI to develop the hardware. Um so we'll we'll talk more about it. Um [1:09] so um if you if you look around you the the daily objects that you use your phone your laptop [1:14] the chair you're sitting on the carpet beneath you everything every hardware every physical [1:20] product has to go through a very rigorous product development process and uh it takes a lot of time. [1:28] the phone that you're holding uh the design might have started like three years or four years back [1:33] um and it took like those many years just to hit um the consumers. So the whole process takes [1:40] significant amount of time right now in developing the that hardware and that's primarily because [1:46] you have to develop the product um design the product test the prototypes [1:50] send it to your contract manufacturers or manufacturing partners that are based [1:54] out sort of like different locations you have to have like contractual agreements with them. [1:58] Um uh you also do a lot of like sourcing related activities um in identifying like where to source [2:03] these components from and um uh you have to do like cost optimization iterate the design based on [2:09] your uh manufacturing uh capabilities. So all of this uh process takes significant amount of time [2:17] and uh with with LLMs with AI becoming more real now we can we can find use cases within [2:24] the hardware development process and try to cut down the time it takes to perform that process. [2:32] Um today I'll be focusing on two different use cases. One is design for X which means [2:38] uh design for manufacturing, design for scalability, design for reliability. um [2:43] all of these recommendations come from different places and the design team has to incorporate all [2:48] those recommendations before this they scale up the product. So we'll we'll um go through that in [2:54] in detail a little bit more um later. Um uh I'll also go through a couple of case studies from [3:01] within Amazon Robotics. Um try to keep it like a little confidential but also try to show some like [3:07] uh the real output that we're getting um in terms of improvements in the timeline. Um and we are [3:13] trying to use AI for the product development process within uh Amazon robotics at different [3:18] programs. So you'll see some case studies based on the real uh improvement that we are seeing. [3:26] Going to the design review process itself it's it's as as I mentioned you have to go through [3:30] significant amount of um significant amount of uh work. Um for example, if you're designing um a [3:39] robot um you you the design team designs it uh and and the whole kind of like the design package is [3:46] created shared it with the manufacturing partners. Uh they could be based out of different places and [3:53] um it contains like 3D uh 3D files, 2D drawings. I don't know how many of you are from like the [3:59] hardware engineering background. Can you raise? Okay, quite a few. Okay. So people um uh yeah uh [4:06] so um we um we we prepare like design package and then send it to our like suppliers. Um usually the [4:13] suppliers look at the design package understand each and every part in detail um and see like [4:20] is this manufacturable? Can we scale this uh robot to like millions or even like more than that? Um [4:29] so the design engineer um has to take all that feedback which is coming from the manufacturers [4:35] and this back and forth takes significant amount of time. Uh for example if I'm talking about like [4:40] the metal parts the die casting there involves like a lot of tooling activities involved. [4:45] You have to uh design the product and then you have to make a lot of changes in your design [4:50] based on the draft angle uh based on the material selection based on the the specific manufacturing [4:56] capabilities at the supplier site. Uh and it differs from supplier to supplier. It depends from [5:02] um manufacturer to manufacturer. Um it depends based on the geographical locations. [5:08] So what what we're trying to do is if we can use AI to catch all those different types of issues [5:15] that we are seeing uh from the manufacturer um after like 8 weeks or 10 weeks if you're able [5:22] to have AI deployed very upstream in the process where we are able to catch all of these different [5:27] issues very upstream in the process that that saves significant amount of time [5:32] um for us to like develop the product faster and deploy the products faster. [5:39] Um I'll give you one quick example of uh how the uh the the feedback from um supplier looks like. [5:47] Uh this is this might be very relevant to people who um who who are who come from [5:52] like hardware engineering background. But this is basically a metal part. It's a sheet metal part. [5:56] Um the designer looks at sorry the the manufacturer looks at this part [6:01] and then they come come back with like several feedback. For example, DFM feedback. [6:05] uh they say okay like the corner here um it has uh the radius which which is uh smaller than expected [6:13] so you have to increase the radius if you don't do it then you'll end up like cracking the material [6:17] over the time or like over time when you when you deploy uh similarly the diameter here is like not [6:23] uh diameter of the hole it's not um matching with the material thickness so you have to [6:28] like increase um some recommendations around like cost reduction if you can consolidate [6:33] all the different holes on a uh you can save on tooling. You can also save on um [6:38] different uh like manufacturing u equipment and you don't have to like buy any equipment for uh [6:45] this to manufacture this part and also you don't have to buy like tooling for this particular part. [6:50] Um and for maintainability uh you can see there's like a small hole here. When people [6:55] try to when when the operators try to fit in all these different components together they have to [7:00] reach two different places within the assembly process. Um and this kind of feedback design [7:05] for assembly design for maintainability is also given by the manufacturers and and the current [7:11] process is basically human reviews the 3D drawings the 2D um parts that that that was sent by the [7:18] person and we're not only talking about like a few parts if if you're deploying like a huge [7:24] robotic solution it has thousands of parts I'm talking about like unique thousands of parts [7:29] um and when you scale it up with multiple programs it it becomes enor enormously high [7:36] uh at the contract manufacturer's site for them to like go through and [7:40] they come back usually like after weeks of um uh after weeks of um analysis. [7:50] So um how how is this possible that we can use AI to compress this timeline? Um [7:58] primarily because of two things. one with uh MCP coming up in like late 2024 um we were [8:06] able to connect our AI with the solid works uh which is which is the CA like native CAD tool [8:13] um and by connecting that we established a bridge between like what is is kind of [8:20] um um like what AI tools can do uh by creating that bridge and by [8:26] by uh by that we were able to extract the material properties the the the features the bounding box [8:34] um by by creating that bridge. So um now you have like information on how the part is kind of uh [8:41] like what what are the different features that are on the part. [8:45] Um many people might have heard the debate between like MCP and API. Um I'll quickly try to address [8:51] this. So we're using they're like people who are using APIs quite a bit but at the same time you [8:56] have to have like defined endpoints. what are the parameters that go into this uh CAD model? [9:00] What are the parameters that are coming out? Um, but with MCP, it's been great because um [9:07] when you connect it with with an MCP, you're able the the agent is able to figure out um what tasks [9:15] uh that the user is asking for and uh define that task and build the code right on the fly which is [9:21] not possible with like hardcoded uh APIs before. The second thing is multimodal analysis and um [9:29] this session like fits right into it. Um not only we are getting feature data, the material data [9:35] from the part, we are also trying to get the screenshots from the part which basically um [9:43] helps AI to get a visual understanding of the part. So we take like multiple screenshots on [9:49] the top view, side view, all the different views. And now you have two different I would say three [9:55] different kinds of data. One you have the feature data which is coming from the MCP and also you [10:00] have the screenshots which is coming um uh as as a part of data collection. Now you feed both of this [10:06] data to an LLM and then ask it to uh perform um a DFM analysis um against like a rule checker. So [10:15] this is only possible now. Uh there are like a lot of rules based engines that that are present in [10:21] the past. But um having a multimodal AI you you're uh understanding how the part looks like. It's [10:28] it's a visual understanding of the part and AI can act as a senior like a seasoned um manufacturing [10:35] engineer who is looking at the part in in real time and understanding how it is built not just [10:40] understanding the feature data. So those are the two things which kind of um are enabling this um [10:49] the the product development process to to be to be faster. Um a little uh background on like what [10:57] happens in the background you basically create like multiple markdown files and um how do we [11:02] build the markdown files? It's basically once you create an MCP you chat with uh your solid works [11:08] um CAD platform and then say like are you able to see a part which is open and then uh it responds [11:14] like okay yeah I I can see the part um but at the same time when you ask it to like extract all [11:19] the feature data it says like I I cannot do that because I don't have like respective tools within [11:24] the MCP um then it tries to create its own tool like if you push a little further uh it tries to [11:30] create it create its own macros VBS scripts on the I and this is where the MCP is very helpful. [11:36] Um it creates like all the things on the fly. It grabs all that data and um I'm like okay can [11:43] you capture the screenshot? It takes a screenshot. Now I'm like can you do all the different analysis [11:48] um of of DFM compared with like the rule checker and whatnot. So it does all of that and at the [11:54] end of the session like the cloud code session I would be like okay can you create a uh markdown [11:58] file based on what we discussed right now. uh that kind of sets a process. If I have to go back and [12:06] analyze another part, I'll I would just like probably call the uh call the existing uh [12:13] markdown file and it kind of like orchestrates um everything based on what what we discussed before. [12:19] And we can do this for multiple things. Uh we can do this for um like metals, plastics, [12:24] um electronics, PCBAs. uh just have the same exercise with them and then save everything [12:29] in a markdown file which are basically agents for you and you can create like a orchestrator [12:33] on top of it with orchestrates different and calls like different agents whenever needed. [12:39] Um this also helps in uh doing like um calling deterministic uh code whenever necessary and [12:47] calling like lm only when you have to call like an an LLM. So that markdown file also handles [12:53] that kind of switch between uh deterministic rule-based approach versus the um LLM approach. [13:02] A quick demo um what you see on the side is basically a part on the right I have a cloud code [13:10] um it's not a demo it's it's a video but um I have like cloud code running on the right and u [13:17] I kind of like load the markdown files into the the session and then ask it to [13:23] perform DFM analysis on the part. It starts running. This is at I think 4x or 5x. Um [13:32] but essentially it it creates a task list of what needs what needs to be done. what are [13:36] the different macros that it needs to extract, take the screenshots, describe the component, [13:42] um evaluate it against the DFM guidelines, um which which are kind of like uh written down [13:52] and once it does, it finally generates an HTML report at the end of it. [13:56] Um you can see I'll show more details like in the next slide [14:00] but you can see it flags different issues on the part um based on the the analysis that it did. [14:09] Uh these are the different outputs that we get. Um for example like you have the the screenshots [14:16] and some details about the part and interesting part is basically this component description. [14:21] This component description is only possible uh because of LLMs and it it it's able to [14:27] understand part holistically rather than just like uh rule-based understanding of of the part. [14:33] So that kind of like sets a hardware foundation. Uh and you can you can ask like different [14:38] questions about your design itself. Um this is a summary table at the end that shows like how [14:45] what are the different parameters that are passing that are failing [14:51] and and and and the and the good thing is like you can modify all these rules [14:54] like on the fly based on your specific manufacturer based on your location [14:58] based on internal like Amazon robotics standard you're always uh you h you have [15:03] that leverage to modify these rules like as and when possible and very quickly as well. [15:10] Uh the second part um that I was going to talk about is like how are we doing cost modeling. [15:15] This is mainly helpful because um the suppliers when they come back with like [15:20] the quotes from the costs we usually um kind of like believe that they're right [15:27] but at the same time we have to do some kind of like double check to understand like if they're [15:31] what whatever they're quoting is that is that right or not. Uh so we are also using um [15:36] like multimodal AI approach to understand what a part should cost and uh eventually like scaling [15:42] up to a program level cost like what what it takes to build this whole robot at at a certain volume. [15:51] uh and that's primarily dependent on u a little bit of techn technicality here like if you [15:55] if if you want to uh estimate the cost there's like material cost there's manufacturing cost [16:00] and manufacturing cost is primarily dependent on like the cycle time if you're doing like machining [16:05] and you multiply cycle time with like the shop hourly rate uh that's how you get the [16:09] manufacturing cost so to to estimate that cycle time um and to understand how the part looks like [16:16] that's where it becomes uh very important for for an AI to be uh part of the process [16:24] um we don't blindly believe in what AI is saying. We um take like evaluation metrics very [16:31] seriously. In this case, since I talked about DFM design for manufacturing and should cause [16:36] um DFM is evaluated based on DFM coverage. basically uh if there's a part that we send to the [16:43] supplier if they come back with like 100 feedback 100 like items of feedback how much of that is [16:49] um AI capable of catching it so that's the coverage and accuracy is basically how much of the [16:57] um DFM feedback that was caught by AI how much is it accurate um so we evaluate the DFM accuracy and [17:04] DFM um based on uh these two metrics and for short cost uh the cost modeling we compare uh [17:13] the the delta between like the supplier cost um versus what the AI is predicting. [17:23] um couple of case studies. There are a couple of programs that we [17:26] kind of like did a case study um a pilot run on and um uh for the DFM for the DFM part [17:35] um we chose like a few plastic injection molding components. Um and and there were like [17:42] 90 different types of feedback that came from the actual supplier. And when we compared it with [17:46] the sub the the um the the AI comparison um the AI tools predictions [17:53] we found um like about 70% of the issues that were flagged by the supplier were caught by AI. [17:59] Um and we're talking about like this 70% was caught in like 2 hours of work. Um but if [18:06] you if we had like relied on the supplier they would have come back with after like two weeks [18:10] because uh they have to like go through all of this different um analysis on their end. [18:17] Um and within the 70% that we caught um 85% was like accurate. remaining 15% it was not accurate [18:25] but the work is currently going on to improve the accuracy uh which is basically um like fine-tuning [18:31] the model based on the previous uh suppliers data and that that helps us to like uh match [18:38] all of these analysis specific to a supplier uh for short cost where the predictions um from AI [18:45] they were about plus - 25% within the quotes uh that we received from the actual supplier [18:51] and um there's a lot of work going on to bridge this delta as well. [18:57] Right? Uh that's about it. Thank you so much for uh listening in.