Valeo massively uses generative AI to increase its production of code. The French equipment supplier has chosen to turn to Gemini for the majority of its use cases.
In recent years, the specialist in embedded systems in Valeo vehicles has seen its needs in code explode. “Today, we are witnessing two major phenomena. Not only is more software in the car, but this software can now benefit from continuous updates during the vehicle life cycle,” said Cédric Merlin, director of AI at Valeo. A boom from the software brick which results in an explosion in the number of code lines produced every day. In parallel, the development times of a vehicle have shortened, going from 4 to 2 years. Result: an increasing need for code, always more complex and critical. “Making an autonomous system without AI has become impossible,” sums up Cédric Merlin.
To cope with this growing tension on code production, Valeo has chosen generative AI. Of its 9,000 systems and software engineers, nearly 5,000 are now equipped with Assist code, an code assistant based on Google Gemini. But beware, there is no question of replacing the developers. “We need fewer developers per quantity of code, but as we need more code in absolute value, we keep the same number of developers,” says Cédric Merlin. The objective is to “increase” them, allow them to focus on high added value tasks such as design, robustness and validation. AI is only there to speed up development cycles, not to replace experts.
Reduce development cycles
Beyond the simple generation of code, Valeo uses generative AI as a real assistant throughout the development process. “We have set up semi-automated workflows that allow us to strengthen our processes,” said Valeo IA specialist. AI agents are working at different stages: generation of unit tests, code review, detection of potential anomalies. “Beyond shortening the provision cycles, the AI makes it possible to produce code in greater percentage while respecting the safety and quality standards,” says Alexandre Cussignac, technical director of Google Cloud.
The AI even offers automatic corrections that the developer can accept or reject. “The objective is to generate tests and converge more quickly on the validation phases”, details Cédric Merlin. Result: shorter development cycles, more robust codes and an ability to deal with increasingly complex projects, says Valeo who does not share KPI around his earnings thanks to AI.
Low level language, a black point
To feed its development strategy, Valeo does not only put Gemini. “We use a model of models,” explains Cédric Merlin. If Gemini remains the main model for the majority of use cases, the supplier is also based in other areas on Claude, Llama and Mistral. The objective is to systematically select the model available on Vertex AI the most suitable.
Despite these capacities, Valeo comes up against a major challenge: the development of low level code, criticism in the automotive industry. “Generative models are much more efficient on languages like Python, but much less on very specialized on -board code”, recognizes Cédric Merlin. To get around this limit, Valeo works in close collaboration with Google on the fine-tuning of models. The objective is to adapt Gemini to the specifics of the automotive code, drawing it on corpora of low level code specific to the company.
Finally, in parallel, Valeo has a privileged partnership with Google Cloud to test in Early Preview the latest firm technologies, especially in AI. “When you have access to technology earlier, you can quickly have to correct what does not work and adapt the solution to our specific needs,” explains Cédric Merlin. A positioning that offers a certain competitive advantage.
