
Ameet Deshpande
Curium cofounder; Princeton-trained researcher
Curium
At a Glance
- Role
- Cofounder, Curium
- Doctorate
- Princeton University
- Undergraduate degree
- IIT Madras, 2019
- Research
- Language models, AI safety, and generative engine optimization
Studying Language and Learning at IIT Madras
At IIT Madras in Chennai, India, he worked on reinforcement learning and natural language processing before completing his undergraduate degree in 2019. His research at Princeton developed under Karthik Narasimhan, with interests spanning data-efficient algorithms, model architectures, and multilingual language processing.
The work included methods for learning from limited examples and research on toxicity when ChatGPT was assigned a persona. Those projects considered both what a language model could learn and how its behavior could change with the instructions it received.
Building a Way to Measure Visibility in AI Answers
The GEO project addressed a different kind of influence. When an engine writes an answer from several sources, the sources need a way to understand their contribution. The researchers developed measures of visibility and tested changes to content against a benchmark.
Relevant citations, quotations, and statistics were among the approaches examined. The results depended on the subject and the kind of question, making evaluation an important part of the method.
At Curium, he joined fellow researchers Vishvak Murahari and Narasimhan in applying that work to a commercial platform. The company develops systems to measure and optimize how brands appear in AI search.
Teaching Alongside the Research
He also made teaching part of his work at Princeton, including a ChatGPT workshop in February 2023. “I am deeply passionate about teaching and view disseminating knowledge to be as important as the research I do.” The teaching commitment gives another purpose to the papers and tools.
The move from language-model research into GEO brings those interests together. Understanding how an answer is generated helps explain both how a system can be influenced and what information its users and source publishers need to understand about the result.
Links
Sources
Portrait retained from the original profile; image unchanged.