PI: Ahmed Rezine, IDA, Linköping University
co-PI: Amir Aminifar, EIT, Lund University
Agentic applications are revolutionizing our usage of computers but come with unprecedented safety and security challenges. We focus on agentic applications that generate software involving neural networks for the embedded domain. We propose to develop a framework to reason about correctness, to make explicit assumptions, to automatically validate reasoning, and to systematically resort to “runtime guards” for the parts that are out-of-reach of mechanical inference. As a result, users get stronger guarantees when they get help from agents to generate embedded software, they will be able to track the assumptions, to propagate them, and eventually to assign blame. We believe this will represent a strong scientific and industrial contribution for embedded software.
Project number: H8