Karan Sharma

Title: Karan Sharma and Secure Agent Training

Introduction

Karan Sharma is an inventor based in San Francisco, California. His patent work concerns security safeguards for training agent-based machine learning models.

Latest Patents

Sharma is associated with 1 patent, titled “Utilizing a Multi-Layer Security Architecture for Agent Reinforcement Fine-Tuning of a Machine Learning Model.” The patent describes systems and methods for fine-tuning an agent machine learning model through reinforcement learning, using user-defined grader endpoints and external tools.

The disclosed architecture can filter incoming training datasets and conduct content-refusal checks, chain-of-thought leak detection, and governance oversight before, during, and after training. It also describes stateful trajectory rollouts that link training trajectories to unique identifiers, helping multiple environments and tasks run concurrently while maintaining consistency across tool calls and grading interactions.

Career Highlights

Sharma is associated with OpenAI OpCo, LLC. His patent reflects work on combining reinforcement fine-tuning with security controls throughout the model-training process.

Collaborations

The provided information lists John Allard and James Blomo as Sharma’s coworkers. It does not specify their roles in the patent or whether they collaborated on it.

Conclusion

Karan Sharma’s patent focuses on securing agent-model reinforcement fine-tuning through layered safeguards and consistent tracking of training trajectories.