Gangavarapu, Tushaar

Tushaar Gangavarapu


MS CS student at Cornell University
Contact: tg352@cornell.edu (by email), or 456 Gates Hall in person

"[...] But intellectual honesty is more than just acknowledging the work of others and not outsourcing your commencement speech to a neural network. It's also about presenting work in an unbiased manner, about making compelling arguments without misleading, not confusing wanting something to be true really, really badly with it actually being true. [...] There's a fascinating and scientifically very significant instance of intellectual honesty in Newton's understanding of gravity."

—Prof. Alyssa Apsel, ECE graduation ceremony, May 2023

I am an MS CS student at Cornell University (Ithaca), advised by Alexander "Sasha" Rush. My work is at the intersection of alternate-attention (for large language models), ML systems, and mechanistic interpretability:

        Alternate-attention: Compute/memory-efficient architectures such as Mamba(Byte), RecurrentGemma
        :: ML systems: Hardware-aware kernels (in Triton)
        :: Mechanistic interpretability: Sparse autoencoders to understand the role of recurrence in linear recurrence models

Please see my GitHub and research pages to learn more about my current projects. I no longer actively work on pre-2022 research noted on my research page (e.g., healthcare analytics). To this end, I believe there are more qualified experts in those areas, and as such, I may not be able to respond if you are reaching out regarding extensions of that earlier work.

If you're a Cornell undergrad and interested in any of my work (or want to discuss CS4740 or related topics), feel free to reach out! I'm generally proactive with emails and happy to talk about (my) research.


Research interests. NLP :: LLM :: alternate-attention :: efficiency, mechanistic interpretability

Recognition. For my contributions to CS4740 (and cross-listings) in fall 2022 and 2023, and CS4300 (and cross-listings) in spring 2022 and 2023, I was awarded the Bowers CIS Best TA Award.


Timeline.

08/2024–01/2025: Head TA, Practicum in AI (CS4701) Fa24, Cornell University, Ithaca
05/2024–08/2024: Research Intern (Advisor: Sasha Rush, Project: Mechanistic interpretability), Cornell Tech, NYC
01/2023–05/2023: Head TA, Language and Information (CS/INFO4300) Sp24, Cornell University, Ithaca
08/2023–01/2024: Head TA, Natural Language Processing (CS4740/LING4744/COGST4740/CS5740) Fa23, Cornell University, Ithaca
05/2023–08/2023: Research Intern (Advisor: Sasha Rush, Project: Authorship identification), Cornell Tech, NYC
01/2023–05/2023: Head TA, Language and Information (CS4300/INFO4300) Sp23, Cornell University, Ithaca
08/2022–01/2023: Grad TA, Natural Language Processing (CS4740/LING4744/COGST4740/CS5740) Fa22, Cornell University, Ithaca
08/2022–now: MS CS Student (Major: Computer Science, Minor: Applied Math), Cornell University, Ithaca
09/2021–08/2022: Applied Scientist, Kindle Content Experience and Quality Algorithms, Amazon
12/2019–09/2021: Research Engineer, Kindle Content Experience and Quality Algorithms, Amazon
06/2019–12/2019: Software Development Engineer, Worldwide Deals, Consumer Engagement, Amazon
08/2018–12/2019: Scientific Researcher at Human-Centered Computing Group and Healthcare Analytics and Language Engineering Lab, NITK Surathkal
05/2018–07/2018: Software Development Intern, Kindle Create, Publisher Tools to Reader, Amazon
05/2017–07/2017: Research Intern, Center for Pattern Recognition and Machine Intelligence, PES University
08/2015–04/2019: Undergraduate Student (Major: Information Technology, Specialization: Aritificial Intelligence), NITK Surathkal

Misc.

  • I really like chocolate-chip cookies, and hate yellow lays!
  • I am scared (but also excited), a lot!, while teaching, publicly speaking, etc.
  • Lecture notes on backpropagation (personal knowledge + compiled from various sources)
  • Oh!, I love sleeping; when I'm not a computer scientist, I'm probably (read: definitely) asleep


Scholar | GitHub | Twitter

Tushaar Gangavarapu [tg352@cornell.edu];