Zürich AI Meetup | Secure Orchestration

Agenda: 17:30 – 17:45 – Welcome 17:45 – 18:15 – Steffen Hausmann (LangChain): How We Built it: LangSmith Engine 18:15 – 18:45 – Yixuan Xu (ETH AI Center): From Open Model to Open Multimodal: Building Apertus 1.5 18:45 – 19:15 – Edgar Kussberg (Sonar): Now We Keep Agents Honest 19:15 – 19:45 – Max Rumpf (SID): How to Train a Frontier LLM with RL (Auth0): Securing AI Agents — Workshop Steffen Hausmann (LangChain): How We Built it: LangSmith Engine Abstract: Steffen will walk through LangSmith Engine—an agent that analyzes traces, identifies recurring failure patterns, and suggests concrete next steps. He’ll cover its architecture and how teams operationalize the agent-improvement loop. Yixuan Xu (ETH AI Center): From Open Model to Open Multimodal: Building Apertus 1.5 Abstract: Apertus 1.5 extends the open language-model initiative into a fully open multimodal foundation model spanning vision and audio. Its multimodal understanding comes directly from the LLM backbone, trained on discrete multimodal tokens rather than external encoders. The talk covers the team, journey, key ideas, and lessons learned. Edgar Kussberg (Sonar): Now We Keep Agents Honest Abstract: Rebuilding a 15-year-old code-analysis company for the agentic era—and the three things that bit the team along the way. Max Rumpf (SID): How to Train a Frontier LLM with RL Abstract: RL has driven many recent LLM intelligence gains, but its techniques remain poorly understood. Based on training SID-1, an agentic search model, this talk explores the infrastructure, challenges, data, and practical limits of reinforcement learning for frontier LLMs.