NASE (Neural Architecture Search Engine)
Agentic Artificial Intelligence Accelerates Embedded Development
The Fraunhofer ITWM, together with the Rhineland-Palatinate Technical University Kaiserslautern-Landau, is presenting a system that fundamentally accelerates the development of embedded AI applications: »Neural Architecture Search Engine« with Hardware in the Loop.
With NASE (Neural Architecture Search Engine) Fraunhofer ITWM (Institute for Industrial Mathematics) presents an agentic system that designs and optimizes AI models for edge and embedded platforms. Instead of laboriously adapting existing models to a target platform, the neural search develops tailor-made architectures directly for the respective hardware, eliminating the need for additional model porting.
Software Tool for Hardware-Optimized Models in Weeks Instead of Months
»In practice, developers often resort to freely available models that do not function satisfactorily on the target hardware,« explains Dominik Loroch from Fraunhofer ITWM. »The necessary adaptation requires expertise in both hardware and AI design. NASE Hardware in the Loop takes care of precisely this complex optimization process.«
A special innovation is agentic AI: a language model queries the requirements for the application and automatically configures the search to achieve the best results. This eliminates the need for long training periods in tool operation.
The system reduces development times from several months to a few weeks – while delivering higher model quality compared to manual design processes.
NASE is particularly interesting for developers who need to port and optimize AI models on embedded hardware.
Typical areas of application include computer vision, production and condition monitoring, healthcare applications, and industrial assistance systems. Basically, all companies that want to use embedded AI in their products benefit, especially where computing power, energy consumption, and costs are strictly limited.
Demonstrator Shows Complete Pipeline
At the trade fair stand (Hall 4, Booth 422), the demonstrator shows visitors the complete development chain from the initial interaction to the executable application on the target hardware. The agentic system first queries the target variables, such as accuracy or runtime, and assists with the configuration of the search. It then generates models that work measurably faster and more efficiently on the respective platform than typical off-the-shelf models from the Internet.
In this context, »Agentic Artificial Intelligence« means that users simply formulate the goal, for example, what task the AI should perform on the embedded platform and what performance indicators must be achieved. The system decides independently how this goal is to be implemented technically. It uses a language model, calls up external tools, checks intermediate results, and optimizes iteratively until the requirements are met. Developers no longer have to specify how the AI model should be structured in detail; they only define the »what,« not the »how.«
Hardware in the Loop: Real Measurement Data Instead of Estimates
In the embedded sector in particular, the actual performance of an AI model is difficult to predict. That is why »Hardware in the Loop« is so important: models are executed and measured directly on real hardware. The RPTU (Rhineland-Palatinate Technical University Kaiserslautern-Landau) in Kaiserslautern has developed a system for this purpose that can be used to automatically evaluate different hardware platforms. The measurement data obtained flows directly back into the architecture search. This results in models that are not only theoretically, but also demonstrably optimally matched to the target hardware.
STANCE Networking Evening
As part of the trade fair, Fraunhofer ITWM and Fraunhofer IIS will also be hosting the STANCE networking evening at JOSEPHS in Nuremberg on March 11, 2026.
STANCE is an alliance of industry and research founded by Fraunhofer ITWM and Fraunhofer IIS that promotes neuromorphic and spiking-based, particularly energy-efficient, and extremely fast AI technologies for industrial applications. The event offers compact, practical insights into neuromorphic computing and creates space for exchange between industry, science, and start-ups.
