# Road to GTC Berlin: Europe’s AI Builders in the Spotlight

## Executive summary

This video summarizes the 'Road to GTC Berlin' build series across Europe, highlighting key developer themes in AI, particularly agent development, open models, and physical AI. Speakers demonstrated advanced use cases ranging from healthcare workflow automation and patent law analysis to advanced robotics and computer vision. A major recurring theme is the need for agent safety, security, and the deployment of sovereign, local AI solutions (on-prem/edge) to ensure reliability in mission-critical environments.

## Key takeaways

- Agent Safety and Security: A top concern across all cities is making agents safe and secure. Solutions include using OpenShell (containerization) and implementing frameworks like Nemo Guardrail to sandbox agents, ensuring they cannot execute unauthorized actions even if they possess the capability (e.g., preventing an agent from purchasing items via an API call when only reading is allowed).
- Sovereign and Edge AI Deployment: For safety-critical fields (like healthcare), relying on external cloud models is risky due to fluctuating pricing and stability. The trend is toward sovereign AI, enabling local, on-prem deployment using hardware like the NVIDIA DGX Spark or RTX Spark to maintain full control and data privacy.
- Multimodal and Computer Vision Agents: Advanced agents are moving beyond simple code or API calls. Demonstrations showed agents using multimodal capabilities to 'see' a screen (GUI) and interact with applications (e.g., CAD software, ERP systems) as if a human were operating the computer, enabling complex, natural language-driven workflows.
- Physical AI and Robotics: Robotics applications are advancing rapidly, moving from simple tasks to complex, agentic workflows. Examples include training robots for bin picking using imitation learning and reinforcement learning, and developing voice-driven AI assistants for hospital staff to keep their hands free for patient care.

## Technical details

- Agent Architecture & Safety: The concept of building agents involves more than just using services (like ChatGPT). Key components include defining agent ingredients, using containerization (OpenShell), and implementing guardrails (Nemo Guardrail framework) to restrict agent actions.
- Local/Edge AI Hardware: Devices featured for local AI prototyping include the NVIDIA DGX Spark, RTX Spark, and the Jetson family, allowing developers to run models on-site, ensuring data sovereignty and reliable operation.
- Large Language Models (LLMs) & Open Models: The focus is on open-weight models (e.g., Gemma, Neatron 3.5) that can be deployed locally. Techniques include using embedding models (NVIDIA embedding model) and vector databases to process and compare vast amounts of data, such as comparing historical texts (Britannica) to modern Wikipedia.
- Robotics Learning Paradigms: Robots are being trained using advanced methods like imitation learning and reinforcement learning (RL). For instance, moving a robot arm for pick-and-place tasks was shown to improve significantly when RL was applied to the policy stack.

## Practical implications

- For build engineers, the emphasis on local, sovereign AI deployment suggests a shift away from purely cloud-based solutions for sensitive or mission-critical systems.
- The integration of multimodal capabilities (vision, GUI interaction) means that future agent designs must account for interpreting visual interfaces, not just structured APIs.
- The use of reinforcement learning and imitation learning in robotics suggests that complex physical tasks can be accelerated by feeding success examples into RL architectures.
- Implementing robust safety mechanisms (guardrails, sandboxing) is now a core requirement for any agentic workflow handling sensitive data.

## Topics

Artificial Intelligence, Agentic Workflows, Robotics, Edge Computing, Healthcare Technology, Patent Law Automation, NVIDIA DGX Spark / RTX Spark, Nemo Claw, OpenShell, Neatron 3.5, Knowledge Across Time

Source: https://www.youtube.com/watch?v=v_Nc84xiCF4