At dAIEDGE, we continue our Meet the Winners series by highlighting the innovators pushing the boundaries of AI deployment in real-world environments. In this interview, we feature Anil Uzengi, winner of our Second Open Call, and the developer of the project STROMA.
STROMA addresses a critical challenge in deploying advanced AI systems at the Edge: how to make active inference models-widely used in robotics, automation and safety-critical systems, run efficiently under strict hardware constraints. These systems continuously interpret sensor data and decide on the next action, requiring fast and reliable inference and control loops. However, in practice, they are often limited by high memory usage and latency, making real-time execution difficult on edge devices such as NVIDIA Jetson.
The core objective of STROMA is to bridge this gap between theory and deployment. The project focuses on building a JAX-first optimization and deployment pipeline that significantly reduces memory footprint and latency in the most critical parts of the system, the inference and control loops. By applying advanced optimization techniques, STROMA enables these models to meet real-time requirements even on resource-constrained hardware.
Beyond optimization, the project also ensures practical deployability. STROMA exports optimized models to ONNX and builds TensorRT engines, enabling efficient execution on NVIDIA Jetson devices. This end-to-end pipeline transforms experimental models into production-ready solutions, capable of operating reliably in real-world scenarios.
During his journey within dAIEDGE, Anil developed his work in collaboration with VERSES, acting as hosting institution. We would like to express our sincere appreciation to VERSES for their support and expertise, particularly in guiding the project towards a strong deployment-oriented approach.
A key part of this journey was the opportunity to bring the project beyond theory. Throughout the seven-month programme, STROMA was progressively tested and refined within the dAIEDGE-VLab, allowing Anil to experiment under realistic constraints in terms of latency, memory and power consumption. This hands-on validation environment played an important role in bridging the gap between research and real-world performance.
Reflecting on the experience, Anil highlights how the programme helped shift the focus from theoretical performance to real-world constraints. Working under strict latency, memory and power limitations provided valuable insight into the challenges of deploying AI systems outside controlled environments, reinforcing the importance of optimization and robustness.
STROMA exemplifies the type of innovation that dAIEDGE aims to foster: solutions that not only advance AI methodologies, but also ensure they are efficient, scalable and ready for deployment in demanding edge environments.
Want to see the full story behind STROMA? Watch the complete interview here.