Historical incidents underscored the urgency of our mission. On February 25, 1991, an American Patriot missile deviated 600 meters off course due to a minor internal clock counter error. This resulted in the unfortunate death of 28 American soldiers in Dhahran, Saudi Arabia. Another mishap occurred on June 4, 1996, with the Ariane 5 flight 501. A mere 39 seconds after lift-off, the rocket exploded due to a computer failure: the conversion of a 64-bit floating-point number to a 16-bit integer exceeded the latter’s capacity.

Such industrial accidents illuminated the need for improved validation in computer systems. The industry, recognizing the gravity of such errors, sought research-based solutions to prevent and rectify them.

After two decades of dedicated research, Numalis has honed an intricate skillset that amalgamates static code analysis via abstract interpretation, code synthesis, and statistical analysis.

Then, with the support of the DGA, Numalis was able to extend its activities to the validation of artificial intelligence calculations, and more specifically to proving the robustness and explainability of neural networks. The Saimple tool is the result of this work.

While the products we develop may have changed, our mission to provide a level of confidence and accuracy to our customers has remained the same.

Recognizing the importance of standardization to make AI more robust, Numalis is working with the AI community to develop standards. The ability to prove the reliability of artificial intelligence can help highly regulated industries (such as transport or healthcare) to overcome the regulatory and certification hurdles posed by these technologies. To this end, Numalis is developing new standards within the ISO/IEC and CEN CENELEC committees to define new AI validation processes.

We’re hiring