Embedded Austria No. 7: Embedded AI
About this Event
Embedded Austria No. 7: Embedded AI
Expect concrete insights, honest lessons learned, and discussion beyond marketing slides.
The evening concludes with open networking, an opportunity to continue the discussion and connect with others from the embedded community.
Agenda
🕑: 06:00 PM - 06:20 PM
Event Opening
Host: Bernhard Trinnes | Embedded Austria
Info: Embedded Austria is a platform to connect experts, exchange ideas, and learn about the future of embedded systems. Get to know about Embedded Austria and its plans for the future.
🕑: 06:30 PM - 07:15 PM
AI-as-a-Service in Space for Real-Time Earth Observation Intelligence
Host: Dr. Thomas Pusztai
Info: Thomas Pusztai obtained his PhD in computer science at the Distributed Systems Group of TU Wien. His research focuses on the Edge-Cloud-Space computing continuum. As co-founder and CEO of LeoTrek AI GmbH he is working on a software platform for distributed AI-as-a-Service in space to deliver intelligence for time-critical use cases.
🕑: 07:05 PM - 07:45 PM
Edge Learning: Privacy by Design or New Attack Surface
Host: Rudolf Mayer
Info: Rudolf Mayer is a senior researcher at SBA Research, and leads the Machine Learning and Data Management team. He is also teaching at TU Wien.
He has been working in machine learning for 20 years, and tackles privacy and security in ML for the last 10 years, with a focus on federated learning, integrity and confidentiality attacks on ML, and IP Protection of ML assets like datasets and models.
At SBA Research, his research focus lies on machine learning, specifically including privacy-preserving learning and security of machine learning, and data management, among others in the currently ongoing research projects FeatureCloud (EU H2020), focusing on privacy-preserving federated learning, and projects funded by the Austrian Research Promotion Agency (FFG): WellFort, KnoP-2D, PRIMAL, and Gastric.
🕑: 07:50 PM - 08:30 PM
LLM Inference Is a Memory Problem
Host: Jonas Vander
Info: Fixed slots, paged pools and cache side channels, from edge devices to servers
Same language model, same GPU, and yet one inference engine delivers 19 times the throughput of another. The reason is not compute power, but memory management. From an embedded perspective, this talk shows why the KV cache determines how many users and how much context a machine can handle, why fixed slots and aggressive quantisation as in llama.cpp are a good fit for a single device, and why a shared prompt cache can leak information through timing.
Jonas Vander is co-founder and CTO of Xinity, a Vienna-based sovereign AI infrastructure startup, and has spent a decade in AI engineering since studying at Radboud University. He is the architect behind the Xinity Runtime, the open-source platform that lets organisations run AI models on their own servers.
🕑: 08:30 PM - 10:00 PM
Open Networking
Where is it happening?
Event Location & Nearby Stays:
EUR 10.00
