Using AI to automatically evaluate and manage specifications for automotive components

The automotive industry is facing growing challenges in the evaluation of specifications and requirements for automotive parts. Traditional manual evaluation methods are time-consuming and labour-intensive. The use of LLMs now offers an efficient solution to automatically analyse specifications and find potential problems or inconsistencies. This can save time and resources while improving the quality and accuracy of the assessment.

LLMs to optimise a process without loss of quality

Automotive suppliers are faced with a central problem: the enormous number of specifications to be evaluated. As highly specialised engineers and product managers have to review thousands of specifications per month, this results in high costs and time delays in production processes.

Automating this process is extremely difficult, as the technical language used in the specifications presents a hurdle. For example, it is often unclear to laypeople what exactly is meant by requirements such as "The component integration must have a minimum specification of 0.001 micrometres for surface roughness and guarantee thermal stability from -40 °C to 120 °C at a rate of 100 °C per second".

Conventional automated approaches, such as rule-based systems, are not able to deliver the required precision and contextualisation. Faced with these challenges, the use of Large Language Models appears to be a promising solution, as LLMs have made dramatic improvements in natural language understanding in recent years and can learn the correct assessment based on past data.

Analysis platform with individually trained language model

The automotive supplier has developed a platform that can be used throughout the company to evaluate and analyse the specifications of digitised requirement specifications. Due to the Group's international orientation, locations worldwide benefit from the solution.
The core of the platform is a web application based on a microservice architecture. The application is connected via API interfaces to a database in which the specifications are digitally managed. The microservices use a language model that categorises the requirements in the specifications into classes, recognises changes in the wording and links applicable documents.
The language model is based on Google's LLM BERT (Bidirectional Encoder Representations from Transformers) and was retrained on specifications that had already been evaluated. In this way, the model learnt subject-specific vocabulary and evaluation criteria according to which it can classify the requirements.

Product management workload reduced by a factor of 20 to 50

The use of LLM has significantly optimised the evaluation of specifications. The manual evaluation by requirements managers and product teams took 21 to 30 days, whereas the AI only needs one day.
The costs associated with the process were also reduced many times over. The monthly cost for the full manual evaluation of multiple requirements specifications is around €50,000, while the automated evaluation by the AI is around €1,000 per month, including support and maintenance.
Results showed that the AI achieved a high level of accuracy in classification, in some cases even surpassing the human assessors.

Our services

requirements engineering
agile software development
Creation of proof-of-concepts
Backend development
Further development and evaluation of language models
Supporting the machine learning life cycle from problem description to model monitoring
Development of automation modules
Creation of data pre-processing and training pipelines

Used technologies

Python (Pandas, Huggingface, PyTorch)
REST, FastAPI, Swagger
Azure DevOps
Azure ML
Azure Kubernetes Service (AKS)

Modular expansion of features

Once the web platform had been developed, the services could be called up via API endpoints. The existing features were customised to the needs of the product teams in close collaboration with the automotive supplier's product management.

By implementing them as microservices, new features can also be added without jeopardising the productivity and maintainability of the existing ones. A division into microservices also makes sense for the development process because the features are developed in independent teams. As a result, the platform gradually gains more mature functions, which are used to analyse and evaluate a large number of specifications.

Contact

So what can we do for you?

AI-based classification