Reward Function As A Service: A Relatively Easy Recipe for Training Your Own Reasoning Model

This conference talk from MLOps World: Machine Learning in Production provides a 35-minute walkthrough of a template for post-training Deepseek-style Large Language Models using custom reward functions. Learn from Ville Tuulos, CEO and Co-founder of Outerbounds, who brings over two decades of ML infrastructure experience including leading Netflix's ML infrastructure team that created Metaflow. Discover how to customize high-performing LLMs for specific domains using programmatic evaluation through simple Python functions rather than traditional supervised fine-tuning. While this technique is commonly applied to math and coding, the presentation demonstrates creative applications across many other domains. Gain practical insights from an experienced ML infrastructure leader on implementing your own reasoning models with custom reward functions.