From production LLM Logs to better models

Production AI systems generate enormous amounts of valuable data: user prompts, model responses, feedback signals, and interaction patterns.
But most teams never use it.

Inference logs often sit in storage, disconnected from the training pipeline. As a result, improving models becomes slow, manual, and disconnected from real product behaviour.

In this session, we’ll show how teams turn production LLM logs into structured training datasets and use them to improve models through post-training.
You’ll see how to build a continuous improvement loop:
production inference → dataset → post-training → improved model → deployment
This is how AI teams move from experimentation to production systems that get better over time.

What You’ll Learn

• How to capture and explore production LLM inference logs
• How to identify useful training examples from real usage
• How to transform logs into structured training datasets
• How to run post-training workflows on Nebius Token Factory
• How to deploy improved models back into production
• How to build a continuous improvement loop for LLM systems

Who Should Attend

• ML engineers improving model performance with real usage data
• AI developers building copilots, assistants, and AI applications
• Platform teams designing data and training pipelines
• Founders and product leaders building AI-native products