Tusher Chakraborty

概要

I am a Researcher at Microsoft Copilot Tuning, where I work on aligning the LLMs underlying Microsoft Copilot with enterprise customer preferences. My work sits at the intersection of reinforcement learning, reward modeling, and LLM post-training, with a focus on making alignment both technically rigorous and practically scalable.

A central challenge in this space is that comprehensive human annotation is expensive and hard to scale. My research addresses this from multiple angles: designing hybrid human-AI feedback pipelines that concentrate human effort where it matters most, building reward signals that capture reasoning quality at the token level without succumbing to reward hacking, and developing rubric learning systems that continuously discover new quality dimensions rather than saturating on a fixed set of criteria. More recently, I have extended these alignment techniques to long-horizon agentic workflows, where Copilot must handle multi-step reasoning and tool use across live enterprise environments.

My journey at Microsoft commenced with a remarkable project, FarmBeats: the fusion of AI and IoT to revolutionize data-driven agriculture. A key innovation in this endeavor was the creation of an AI-powered IoT sensor fault detection system, which has become a vital part of the Dependable IoT project. These projects have been successfully integrated into the Microsoft Azure ecosystem, leaving a lasting impact on our technology landscape. I was fortunate to be at the forefront of a transformative initiative—enabling IoT communication within the uncharted territory of TV White Space (TVWS) spectrum. This work led to the U.S. Federal Communications Commission (FCC) adopting regulations that allow IoT devices to operate in unused TV channels for data transmission. I also worked on the development of real-time, global insights through the combined power of satellite-based AI and IoT. This work played a pivotal role in the launch of Microsoft’s first research satellite.