The One About Supervision & Guardrails (ft. OCBC AI Lab and Dr Koh Pang Wei)
More information to come soon, keep a lookout! 👀
As language models tackle increasingly intricate tasks, the challenge of oversight shifts from simple verification to complex supervision. How do we provide high-quality training signals when the tasks themselves outpace our ability to verify them at scale?
More About the Sharings
Abhishek Sharma (Machine Learning Engineer, OCBC) will be sharing on "Building an LLM Real-Time Guardrails Service"
Implementing guardrails for LLMs in production environments presents unique technical challenges. Abhishek will walk through their experience building their own real-time guardrails service, highlighting the real-world challenges his team faced including latency constraints, policy flexibility requirements, and CPU-only model deployment limitations. Learn more about concrete engineering steps and architectural decisions they made to overcome these obstacles while maintaining production reliability. (Technical Level: 200)
Dr Koh Pang Wei (Assistant Professor, University of Washington) will be sharing on "Finding Supervision for Complex Tasks"
As LLMs transition to tasks that require significant expertise to verify, acquiring training data at scale has become a primary bottleneck. Dr Koh Pang Wei will discuss three distinct research approaches to scaling supervision:
Delta Learning: Exploring how models can learn effectively from relative quality differences in paired data, even when the data itself is of lower quality than the model’s own output.
RLVE (RL with Verifiable Environments): Using verifiable environments to adaptively generate training problems calibrated to the model’s capability—ensuring they are neither too simple nor too difficult for efficient reinforcement learning.
DR Tulu: Insights into training long-form deep research models using RL with evolving rubrics, providing precise discriminative signalling even for the most complex research tasks.
More About the Speakers
Abhishek Sharma is a Machine Learning Engineer at OCBC's AI Lab, where he specialises in building intelligent software systems. With experience spanning both industry and academia, he focuses on machine learning engineering and operations, including Ray/FastAPI serving, LLM guardrails, and natural language processing. Abhishek brings expertise in developing production-ready ML systems and ensuring robust deployment of language models in enterprise environments.
Pang Wei Koh is an assistant professor in the Allen School of Computer Science and Engineering at the University of Washington, a research scientist at the Allen Institute for AI, and a Singapore AI Visiting Professor. His research interests are in the theory and practice of building reliable machine learning systems. His research has been published in Nature and Cell, featured in The New York Times and The Washington Post, and recognized by the AI2050 Early Career Fellowship, MIT Tech Review Innovators Under 35 Asia Pacific award, Google ML and Systems Junior Faculty Award, and best paper awards at ICML, KDD, and ACL. He received his PhD and BS in Computer Science from Stanford University. Prior to his PhD, he was the 3rd employee and Director of Partnerships at Coursera.
More About the Series
AI Wednesdays is Lorong AI’s weekly gathering, bringing together practitioners, researchers and innovators for technical discussions on research insights, product development and engineering practices.
Get involved: Learn more about Lorong AI | Speaker Sign-up | WhatsApp Community | LinkedIn | X