Better development with AI: a look back at a concrete experiment

This realisation naturally led to SDD, which structures development around solid specifications before utilising AI agents.

“The important thing is to contextualise AI in order to obtain deterministic results, relying on standards and documentation that enable code agents to be consistent with the existing base, while keeping an eye on product security and quality,” Julien Landuré, CTO at Tech Town.

SDD experimentation at SNCF Connect & Tech

The experimentation lasted five weeks, divided into two phases: two weeks dedicated to product and specifications, and three weeks to implementation. During the first two weeks, a product owner/developer duo worked side by side, isolated, to refine the product vision and build useful documentation before any code generation.

This atypical approach favors monitoring, tool evaluation and iteration, but how do you do this when the ecosystem is in perpetual motion? For Emmanuel Cordente: "In a world where every new product is presented as revolutionary, you have to test it to really understand what it can contribute."

Continuing experimentation

Although tangible gains have already been made, the project has yet to be validated for widespread use. The experiment has shown that it is possible to accelerate delivery while producing a high-quality documentary heritage, useful for on-boarding new collaborators and improving testing and maintenance. However, the adaptability of the model to other projects and to the specific needs of different teams needs to be confirmed.

"The long-term goal is to move these approaches from the IA program to standard engineering practice. This ambition marks the transition from an isolated test to a lasting transformation of practices", Emmanuel Cordente.