Python RAG-AI Chatbot for CAD Design Reuse

I designed and developed an AI-powered platform that helps semiconductor engineers discover, understand, and reuse Intellectual Property (IP) from previous projects.

The platform combines Retrieval-Augmented Generation (RAG), multi-agent AI, semantic search, and structured data retrieval to make complex CAD design knowledge accessible through natural language. Engineers can describe their technical requirements, discover matching designs, explore their characteristics and dependencies, and import selected IP directly into their Cadence environment.

The goal is simple: help engineers build on existing, proven designs instead of starting from scratch, reducing duplicated work, accelerating development, and making better use of the company's technical knowledge.

Project context - Reducing CAD Design Cycle

One of the biggest challenges in semiconductor design is that valuable IP already exists across previous projects, but finding and reusing it isn't always straightforward. Design data is often stored in specialized CAD formats, making it difficult to search using traditional methods. Even when a suitable design exists, engineers may spend considerable time finding it, understanding its specifications, and integrating it into their own environment.

I'm working on an AI-powered solution that addresses this challenge through three main components.

IP Mining and Knowledge Extraction

The first challenge is transforming complex CAD design files into information that an AI system can understand and search. I developed a cad-to-text algorithm to analyze design data, extract technical characteristics such as input/output voltages and technology constraints, and generate human-readable documentation describing each design.

I also developed the RAG ingestion pipeline, including document chunking with overlap, embedding generation, and vector storage. This information is stored in PostgreSQL using pgvector, with structured technical characteristics maintained separately to support precise filtering.

The result is a searchable knowledge base that combines the flexibility of semantic search with the accuracy of structured engineering data.

Multi-Agent RAG Chatbot

Once the design knowledge is indexed, engineers can interact with it through a conversational AI assistant. For example, an engineer might ask:

"Find me an LDO with a 3.3V output, compatible with SMOS10HV technology."

To handle these requests, we built a multi-agent retrieval workflow:

• Query Understanding Agent: Interprets the user's request and extracts technical constraints, such as technology=smos10hv and output_voltage=3.3V.
• Retrieval Agent (RAG and Criterias Database): Searches the structured database for matching designs and performs advanced retrieval using hybrid search, context-window expansion, and reranking to identify the most relevant results.
• Response Generation Agent: Organizes the retrieved information into useful answers, including design descriptions, technical characteristics, hierarchical relationships, parent/child IP, and associated testbenches.

I've worked on the backend services, retrieval logic, REST APIs, and integration of these AI capabilities into the application's user interface.

Secure IP Import and Cadence Integration

Finding a suitable design is only half the problem. Engineers also need a simple and secure way to reuse it. To complete the workflow, I developed backend services that allow users to import selected IP directly from the chatbot into their CAD environment.

After identifying a suitable design, the engineer clicks a "Bring to My Environment" button. Behind the scenes, a secure token-based workflow coordinates the data transfer through a dedicated copy server. A local daemon then registers the imported design within the user's active Cadence session, allowing it to be opened directly in the design tool. This connects the AI assistant to a real engineering workflow, turning a conversational search result into something the engineer can immediately work with.

I developed the REST APIs, secure authentication mechanisms, backend orchestration, and Cadence integration using SKILL.

• I developed in Python / FastAPI the Copy Server allowing secure CAD design download through token system,
• I developed the download pipeline automation from webinterface to user area, including also copy server and secure comunications,
• I worked also on the CAD design files mining to generate human readable datas.

• Transformed specialized CAD design data into searchable documentation and structured technical information, allowing engineers to discover IP through natural language.
• Contributed to indexing more than 10,000 design components, creating a foundation for expanding AI-powered design reuse across the organization.
• Connected conversational AI directly to Cadence, enabling engineers to move from discovering a design to importing it into their workspace through a streamlined workflow.
• Reduce duplicated development effort and supports shorter design cycles. We estimated with end users every time they click on bring to my environment button, it avoid them 3-5d of designing the same.