CAD Technical Documentation Generation in Python

I developed an AI-powered solution that automatically generates technical documentation for semiconductor designs, helping engineers spend less time writing specifications and more time on actual design work.

Originally developed as part of a larger RAG platform for reusing existing Intellectual Property (IP), the documentation generator quickly became a valuable standalone feature. The application analyzes complex CAD design data, extracts technical characteristics, and uses LLMs to produce structured, human-readable documentation aligned with company standards.

Working closely with CAD engineers, we refined the solution over four development sprints, introducing DITA-compatible output and customization options based on real user feedback.

The result is a tool that saves engineers approximately one working day per use, while also improving the quality of the technical knowledge available to our AI-powered IP reuse platform.

Project Context - From a Sprint Demo Insight to a Real Business Opportunity

This project started as part of a broader initiative to help semiconductor engineers find and reuse existing designs through an AI-powered RAG assistant.

One of the technical challenges was making specialized CAD data understandable to an LLM. Semiconductor designs contain complex files and technical information that cannot simply be indexed and searched like regular documents. To address this, I worked on a design-mining solution capable of analyzing CAD data, extracting technical characteristics such as input/output voltages and technology information, and generating human-readable descriptions of existing designs.

Initially, the generated documentation was intended to improve our RAG knowledge base. However, during one of our biweekly demonstrations with CAD engineers, we realized that the documentation generation itself could solve another important problem.

Engineers were spending significant time manually preparing technical specifications, even though much of the information could already be extracted from their design data. We identified an opportunity to turn an internal AI capability into a standalone productivity tool.

Building the AI Documentation Generator

I worked on developing the backend services and AI processing pipeline that transform complex engineering data into structured technical documentation. The workflow involves several steps:

• Design Data Analysis: Parse CAD data and extract relevant technical characteristics, including electrical specifications or technology-related information.
• Information Structuring: Organize the extracted information into a consistent structure that can be processed by an LLM.
• AI-Powered Documentation: Use LLMs through AWS Bedrock to transform technical data into clear, readable documentation.
• Document Formatting: Generate structured documentation aligned with internal company standards, including DITA-compatible output.
• Customization: Provide configuration and tuning options so engineers can adapt the generated documentation to their needs.

My work focused on implementing the extraction logic, AI integration, and Python backend services required to support the generation workflow. An important part of the challenge was making sure that the generated documentation was not only readable, but also technically useful and compatible with existing engineering practices.

Iterative Development with CAD Engineers

After identifying the opportunity during a biweekly RAG demonstration, we prioritized the documentation generator and focused the following four development sprints on turning it into a usable solution. We organized frequent demonstrations with CAD engineers to collect feedback and continuously refine the application. This helped us identify several important requirements:

• Generated documents needed to follow existing company documentation standards (content and format).
• The output had to be compatible with DITA-based documentation workflows.
• Engineers needed options to customize and fine-tune the generated content (optionnal content depending on project context).

We used this feedback to progressively improve the output structure, generation process, and customization capabilities.

Integration with the Existing AI Ecosystem

Another advantage of this project is that the generated documentation serves two purposes.

First, engineers can use it directly to prepare technical specifications without having to write everything manually. Second, the same documentation can feed our RAG ingestion pipeline when we mine an older project, where it is processed through chunking, embedding generation, and vector storage. This improves the quality of our chatbot output because the indexed documenntation is generated by an algorithm we validated with our end-users.

I worked in the documentation generation and characteristics extraction I developed a python pipeline to :

• Extract raw data from CAD files,
• Extract determined list of characteritics,
• Generate technical documentation based on determined chapters and expected content.

• Saved approximately one working day per use by automating a significant part of the technical documentation process
• Quickly turned an internal AI capability already developed into a standalone feature,
• Aligned AI-generated documentation with enterprise standards including DITA-compatible output and customization capabilities to support existing documentation processes.
• Improved the quality of the RAG knowledge base: Generated structured technical descriptions that can be indexed and retrieved, making existing semiconductor designs easier to discover and understand.