AI Innovators Munich with Weights & Biases × JetBrains

Expect technical depth, live demos, and practical lessons learned from real-world teams.

• Experiment tracking & evaluation at scale
• Building, debugging, and evaluating LLM applications
• AI developer tools that improve velocity and reliability
• Collaboration between ML and software engineering teams
• From research notebooks to production systems

• Machine Learning Engineers
• AI / LLM Application Developers
• MLOps & Platform Engineers
• Software Engineers building AI features
• Developer Tooling & Platform Teams
• Anyone experimenting with LLMs in production

Are general-purpose LLMs falling short of your company’s highly specialized practical requirements? While Supervised Fine-Tuning (SFT) is an option, what happens when you simply don’t have enough data?

“On-the-job” Reinforcement Learning (RL) is emerging as the key to filling this gap, enabling models to acquire advanced reasoning and align with highly specific business intents. However, the barrier to entry is notoriously high. Manually comparing GPU providers, building deployment scripts, and configuring infrastructure can delay RL training jobs by hours or even days.

Join us for a deep dive into Serverless RL, a new backend powered by Weights & Biases and CoreWeave designed to abstract away infrastructure headaches.

In this talk, we will cover:

• The Evolution of RL in LLMs: Why “On-the-job RL” is the next frontier for practical AI development.
• Bypassing the Infra Bottleneck: How a serverless backend drops cold start latency from 114 seconds to just 3 seconds while reducing overall training costs and wall-clock time compared to local H100 setups.
• Automating Rewards with RULER: Learn how to implement Relative Universal LLM-Elicited Rewards to bypass the need for manual labels and human feedback entirely.
• Real-World Successes: Discover how companies in food delivery and high-stakes finance are using these pipelines to replace slow frontier models with highly specialized, low-latency 8B parameter agents.

Whether you’re looking to build hyper-fast voice agents or specialized internal experts, this session will show you how to empower your software teams to train specialized open-source models on demand, without forcing them to become hardware managers.

The first frameworks for building AI agents primarily targeted Python and JavaScript, leaving the JVM ecosystem behind. But in just a year, the landscape changed dramatically. New JVM-based agentic frameworks such as LangChain4j, Koog, Spring AI, and Embabel have entered the scene.

In this talk, Maria Tigina shares the JetBrains journey from Python-based agents to Kotlin, which ultimately led to the release of Koog. She’ll explore the advantages and trade-offs of building AI agents on the JVM, and look at real examples where strong domain modeling and type safety play a crucial role in making AI agents more stable, interpretable, and production-ready.

Speaker biography
Maria Tigina started her career at JetBrains as an ML Researcher in the AI Agents and Planning research lab, where she studied local agent capabilities and developed benchmarks and evaluation methods for agentic systems. About a year ago, she joined the newly created Koog team and continued her journey as a Software Developer for the AI Agents Framework, also taking on support for the Kotlin MCP sdk. Combining both her research and development expertise, Maria is passionate about building and improving new agents for JetBrains and beyond.