Accelerating Product Research and Design with Generative AI

By Soumen Saha, VP – Capgemini
By Harvinder Bajaj, Director – Capgemini
By Prateek Agrawal, Sr. Partner Solution Architect – AWS
By Mrunal Daftari, Sr. Solution Architect – AWS
By Faisal Oria, Technical Account Manager – AWS

Capgemini

Manufacturers across industries are increasingly seeking to integrate generative AI technologies to stay competitive. However, many face challenges in unlocking the full potential of these solutions. TELme is a direct response to this need, providing a targeted solution that simplifies and accelerates the adoption of generative AI, helping manufacturers innovate faster and achieve greater operational efficiency.

Recognizing the challenges faced by the manufacturing industry and especially product engineering function in adapting to rapid technological advancements and improving operational efficiency, Capgemini came together with AWS to develop and launch “TELme”, a groundbreaking generative AI (GenAI)-powered solution. Designed to empower engineers at a leading manufacturing company, TELme enables efficient search and summarization from a vast repository of structured and unstructured technical content. It’s designed to handle documents such as product specifications, material information, strength & stress testing results, compliance documents, and research papers using natural language queries. By leveraging AWS’s robust cloud infrastructure and cutting-edge AI capabilities, this innovative solution will enhance engineers’ productivity and accelerate knowledge discovery and problem-solving. This new capability will ultimately drive better decision-making, reducing time-to-market and scaling up new talent with secure and streamlined access to a large engineering knowledge base

Designing electronics products with product-specification requirements is a complex task that involves searching for the right components and understanding how they come together to build an efficient, reliable and sustainable product. This search includes finding the components through multiple product manuals, design specifications, scanning through research papers with images, math expressions and tabular data like specifications, measurements, and features. The information generated should have lineage to track them back to the source to validate the completeness. Information could be quickly used in design tools like Computer Aided Design (CAD) for designing products. TELme transforms the day-to-day work of engineers by significantly reducing the time spent searching through multiple documents to find critical information. Instead of manually sifting through complex technical content, technicians can simply ask TELme natural language questions and receive accurate answers spontaneously. By optimizing information retrieval, TELme enhances productivity, accelerates problem-solving, and ultimately shortens the production design lifecycle, helping manufacturers reduce time-to-market and improve operational efficiency.

Manufacturing AI Capability Maturity Model

There are four levels of maturity of an AI application, be it in the manufacturing process or the digital process supporting the manufacturing process.

Figure 1. Manufacturing AI Capability Maturity Model 

Manufacturing Generative AI Platform

TELme is accelerating the journey towards AI maturity by transforming its capabilities to augment human intelligence, delivering pertinent insights driven by user queries. TELme provides recommendations and design considerations, and creates a foundation for delivering autonomous solutions for design, quality review and approval workflows. For example, it can provide a natural language response comparing the different categories of materials and their usage. It can also recommend non-ROHS (Restrictions for Hazardous Substance) compliant parts.

TELme resolves the challenge of inefficient information retrieval by leveraging generative AI and enabling field engineers to quickly access accurate information. It uses AWS generative AI technologies to extract in the relevant information within the context so engineers can utilize it to reduce search time and focus on product design to improve the product research and design lifecycle.

The solution is structured in three layers. The first layer deals with data preprocessing to convert it to a format that is consumable by the foundation models (FMs). The second layer provides access to a large language model (LLM) along with retrieval augmented generation (RAG) infrastructure to generate accurate results based on domain-specific data. The third layer empowers users to interact with generative AI though a powerful UI where they can use the model as it is or can align the data though the UI upload to get desired results.

Intelligent Data Engineering/Pre-Processing

Engineers have to get insights from 5-10 TB of data in various formats. AWS generative AI services help with extracting data from images, audio, and video using Amazon Textract, Amazon Rekognition, and Amazon Transcribe to put them in Amazon Simple Storage Services (S3) buckets. An AWS batch job picks up this data as it’s ingested and applies the chunking policies and converts this data to vector embeddings which are stored in an Amazon OpenSearch database. Different types of data are stored in separate indexes to be used based on the query intent.

Fine Tuning the Model to Understand Your Domain Language

Vector embeddings created in the previous layer are fed to Amazon Bedrock LLM models such as Anthropic’s Claude model. This model is continually retrained with custom data. This enables the model to understand the electronics domain and product vocabulary better. In addition, there is RAG functionality which allows users to augment the latest updated data as needed. Bedrock agents are used to integrate with existing customer systems like ticketing and CRM to take necessary actions. The solution gives customers the ability to bring their own models using Amazon SageMaker in the future if needed.

This tool is deployed using Amazon Elastic Container Service (ECS) containers and AWS Lambda functions, enabling seamless scalability by automatically adjusting resources to help meet varying demands, maintaining efficient and cost-effective performance as the solution grows. Additionally, the tool is designed to be flexible, allowing for the integration and use of various LLMs to suit different processing needs and requirements.

Generative AI Empowered User Interface

The solution’s user interface is integrated with the customer’s existing IAD for authentication. This enables us to inherit the data access controls from existing mechanisms to make sure that engineers asking the questions gets the responses against the data which they are authorized to access.

TELme is available as an application within Microsoft Teams, meant for internal collaboration, and for ease of access for the end users.

The solution enables users to ask questions through the UI based on organizational data by leveraging RAG techniques or the general knowledge of the LLM, providing accurate, contextually relevant responses drawn from both structured and unstructured data sources.

Users can augment the knowledge base by uploading additional documents when needed. User feedback is key to evolving the model and is captured and stored in Amazon Aurora for further processing.

Responsible AI & Security

Security and responsible AI are the concerns we hear most from customers with generative AI solution adoption in production. This solution as the foundation builds on the AWS Well-Architected Principles and responsible AI best practices. Amazon Bedrock Guardrails enhances the safety and compliance of generative AI applications by providing automated safeguards against harmful content, bias, and inappropriate outputs. AWS security services like AWS Identity and Access Management (IAM) maintains the authorization of data while Amazon CloudTrail maintains an audit log of API calls.

For data security, indexes are organized according to classification levels in the Amazon OpenSearch Database, allowing users to receive responses based on their specific permissions.

The solution is exposed to a limited set of internal users (8K) and has delivered impressive results:

• Peak Active Connections: Over 3K simultaneously, demonstrating scalability.
• Unique Users: Our solution has attracted over 2K unique users on a monthly basis, highlighting strong engagement from the select audience.
• Total Queries Processed: The solution has processed over 30K queries in a short duration, showcasing its efficiency.
• Global Adoption: Adoption across 30+ countries, providing broad, low-latency access.

These metrics reflect the solution’s success in its limited exposure, with significant potential for future expansion.

Conclusion

This solution reduces design lifecycle timelines, enables engineers to spend time on more meaningful tasks, and increases productivity. It provides a more intuitive user experience while taking products to market faster. The solution is scalable, is designed to keep future growth in mind, and can handle up to eighty thousand users.

This could be extended to other industries and use cases where generative AI is solving in similar search problems. To learn more about it, contact Capgemini.

.
.

Capgemini – AWS Partner Spotlight

Capgemini is an AWS Premier Tier Services Partner and Managed Cloud Services Provider (MSP) that is at the forefront of innovation to address a breadth of client opportunities across cloud, digital, and platforms.

Contact Capgemini | Partner Overview | AWS Marketplace