Intelligent management of technical projects in the industrial sector

Michelin

The complexity of today’s industrial environment lies not only in the diversity of projects but also in how these projects interact and are shared within a large company like Michelin, a leader in the automotive sector. The issues resulting from the absence of a unified platform are numerous: duplicated efforts, lack of visibility of effective and standardized solutions, the emergence of divergent solutions, and difficulty finding similar projects from other factories that could provide added value.

Aware that this disconnection affected their competitiveness in a rapidly changing environment, Michelin identified the need to develop an intelligent, AI-driven platform to centralize technical knowledge, making it accessible to all factories. They sought not only to solve the problem of fragmented information but also to develop a tool capable of evolving over time, globally scalable, and aligned with their stringent privacy standards. Moreover, the platform needed to be secure, multilingual, and intuitive enough for any user to easily search for information and receive personalized alerts on topics of interest.

Michelin clearly understood their needs and knew the potential offered by artificial intelligence. However, they couldn’t develop this solution alone; Michelin specializes in tyres and mobility, which meant they required a technology partner specialized in AI. After evaluating several providers that could help materialize their vision, Michelin chose Sciling for its expertise in processing complex data.

About the client:

Each Michelin plant manages R&D projects with a degree of autonomy, always subject to approval by central services, using local databases that hinder interplant project connections. This disconnection, along with varied data and differing project storage methods, posed significant challenges. Each plant organized and processed information differently, complicating integration, visibility, and accessibility for other teams.

Added to this were other factors like acronyms, technical jargon, and language barriers based on factory location—obstacles making it impossible to find or even comprehend relevant information about previous projects unless one was directly involved.

“We frequently encountered the same scenario: one of our teams would get stuck with a problem whose solution had already been found months ago at another plant. They didn’t know how to search for this information, making it impossible to ascertain whether a solution already existed. Thus, they resorted to traditional investigation and testing until they discovered the solution, losing valuable time, agility, and opportunities,” explained Germán Arias, Plant Technical Manager & Digital Manufacturing Facilitator at Michelin.

This situation impacted the company’s competitiveness, making resolving it a priority. Germán Arias and Juan Pablo Herreras, Quality Data Analyst at Michelin ULD, recognized the unmanageable volume of data and saw AI as key to resolving their knowledge management issues. They needed a partner capable of addressing their needs, and Sciling emerged as the ideal choice due to its extensive experience in developing and deploying generative AI and natural language processing solutions, offering an agile framework that minimized inherent innovation risks.

Strategic Approach:

Sciling’s research and technical team extensively analyzed available information over several weeks, identifying key project challenges and defining the optimal strategy. During this phase, they explored six different approaches, ensuring each met Michelin’s high privacy and personalization standards. They evaluated various semantic retrieval techniques and agent-based systems, training them with Python libraries and few-shot learning techniques, which effectively handled structured data (numerical and categorical) but required advanced semantic search methods for unstructured, descriptive texts.

A critical challenge was achieving a secure, autonomous implementation without external connections or commercial APIs. Ion Marqués, ML researcher at Sciling, explained: “We evaluated various options to ensure the solution met Michelin’s strict privacy and operational requirements. The need for data control limited our model choices. Fortunately, our previous experience with similar projects led us to select Mistral, which satisfied all project criteria, from deployment to multilingual support, optimizing global collaboration.”

Another significant challenge was integrating diverse data from Michelin factories, varying widely in formats and languages. Initially, the system integrated data from Valladolid and Aranda de Duero, covering 15 users and 250 projects. Later, factories in Cuneo (Italy) and Waterville (Canada) were added, totaling 406 projects. To adapt to Michelin’s data structure, Sciling conducted detailed preprocessing of project databases, enabling the language model to grasp context and terminology. This allowed the system to automatically recognize relevant projects and adapt to Michelin-specific vocabulary, enhancing search accuracy.

Based on this, Sciling implemented an AI agent system integrating function-calling with the LLM, allowing precise execution of predefined tasks. Another specialized agent identified query intent, ensuring accurate responses for numerical filtering, keyword searches, descriptive queries, or multi-criteria requests. Additionally, the multilingual system supported searches and results in local languages such as Spanish, Italian, or English, ensuring intuitive, personalized access to technical information.

Technologies used

Weeks of extensive research and diligent work resulted in Sciling developing an AI system enabling specific searches, organization, and optimization of Michelin’s factory knowledge base.

The PoC was designed to demonstrate technological feasibility without complex, costly developments. In under two months, Sciling successfully presented the PoC internally at Michelin, validating the concept and securing stakeholder support for subsequent project phases.

Thanks to this system, Michelin anticipates increased visibility of innovation processes company-wide, enabling easy consultation and verification of previously opaque, inaccessible information across participating factories. The system leverages existing factory documents, achieving effective results without altering current data-recording practices, thus facilitating deployment by minimizing organizational changes.

“The lack of a shared platform made technical knowledge access difficult between factories. Sciling’s innovative approach brings us closer to overcoming this challenge and establishing a robust shared knowledge base,” commented Juan Pablo Herreras.

Now, Michelin and Sciling look ahead to developing an MVP integrating innovation project data from all 21 Michelin factories, scaling the solution while maintaining simplicity and personalized access to key information.