Designing Open AI Systems that are Trustworthy - From Models to Systems

This 40-minute conference talk by Vincent Caldeira at DevConf.IN 2025 explores the critical transition from standalone AI models to integrated, agentic systems with a focus on trust, transparency, and safety. Dive into practical approaches for designing and operationalizing trustworthy Open AI systems that effectively balance autonomy with accountability. Learn from recent research on risk alignment, model openness, and agentic AI governance while examining a real-world example of an agentic AI system that utilizes multi-agent orchestration and function calling to achieve dynamic, goal-driven behavior. Discover how to implement core safety practices including risk alignment, user alignment, and real-time intervention controls using open source tools and frameworks for safe deployment in environments with high variability. The demonstration, likely conducted on RHEL AI, covers designing AI systems with safety frameworks, orchestrating multi-agent systems for collaboration, implementing function calling for controlled task execution, and establishing operational governance practices to ensure accountability and responsible usage. Access slides and additional resources through the provided link.