About Me
I am a PhD student in Computer Science at the University of Illinois Urbana Champaign (UIUC), working with Prof. Tong Zhang. My research interests include deep learning, natural language processing, and optimization techniques for LLMs.
Education
Doctor of Philosophy in Computer Science
University of Illinois Urbana Champaign (UIUC)
Aug 2024 - Present
Advisor: Prof. Tong Zhang
Bachelor of Technology (Honours) in Electrical Engineering
Indian Institute of Technology Madras (IITM)
Nov 2020 - May 2024
CGPA: 9.22/10
Minor: Physics, Computing
Publications & Patents
SEE-DPO: Self Entropy Enhanced Direct Preference Optimization
Submitted at TMLR 2025
Towards Optimizing the Costs of LLM Usage
Under Review at NAACL 2024
SmaRt: Smart LLM Router for Optimizing Cost and Performance for Document Summarization Tasks
Submitted to US Patent Office
Quality Aware Token Optimization For Reducing Costs of LLM Usage
Submitted to US Patent Office
Prediction of Hydrogen Storage in Metal-Organic Frameworks: A Neural Network Based Approach
Accepted in Results in Surfaces and Interfaces
Topological Data Analysis Enhanced Prediction of Hydrogen Storage in Metal-Organic Frameworks (MOFs)
Submitted and in peer-review to Materials Advances
PeerFL: A Simulator for Peer-to-Peer Federated Learning at Scale
Under Review at IEEE Internet of Things Journal
Research & Professional Experience
Adobe Research
Towards LLM Usage Cost Optimization
Summer 2023 | Bangalore, India
Guide: Dr. Koyel Mukherjee, Dr. Apoorv Saxena, Athrav Tyagi
- Proposed a novel token-aware text paraphraser that reduced the token consumption of GPT based LLMs by 30%.
- Developed a novel pipeline to parallelly select summarization models in a cascade to use, conditioned on the context and the budget constraints resulting in 84.50% reduction in usage cost and a 3.2% increase in the performance.
- Compared the relation between evaluation metrics and human prediction by conducting a division-wide survey.
Nanyang Technological University
Virtual Simulator for Peer-to-Peer Federated Learning
Summer 2021-Summer 2022 | Singapore
Guide: Prof. Anupam Chattopadhyay, Alka Luqman
- Built a fully virtual federated learning simulator capable of handling both central and peer-to-peer settings.
- Adapted NS-3 network simulator to work with Python and interfaced TAP-devices with Docker containers and NS3 to establish an independent virtual network between the containers controlled solely by NS-3 for simulation.
- Conducted real-world experiments to validate and confirm the accuracy and dependability of the simulator's results.
IIT Madras
Prediction of hydrogen Storage
Summer 2023 | Chennai, India
Guide: Prof. Santanu Sarkar, Prof. Chandra Chowdhury
- Implemented novel neural architecture for prediction of hydrogen storage than achieved state of the art performance.
- Worked on generating topologically-persistent images to be a cue to learn the optimal hydrogen storage.
- Achieved a performance gain of 21% on error rate, on the predictions of the model compared to state of the art model.
IIT Madras
Color Restoration for Underwater Images
Fall 2022 | Chennai, India
Guide: Prof. Kaushik Mitra
- Built a unsupervised depth estimation technique using degraded stereo image pairs for subsequent color correction.
- Using these depth estimates, implemented another neural network for estimating the disentangled representation to estimate the underwater image formation parameters. Finally, correcting the degraded images using these parameters.
IIT Madras
Distortion Invariant Reconstruction model for document restoration
Fall 2023, Summer 2024 | Chennai, India
Guide: Prof. A.N. Rajgopalan
- Created a dataset consisting of various types of geometric and photometric distortion for subsequent image correction.
- Developing a novel pipeline for disentangling the distortion and clean image from the degraded image, for the subsequent generation of more synthetic data using these distortion maps and new clean images.
Key Technical Projects
Tool Creation for Large Language Models
Research Project | Spring 2025
- Currently developing a methodology for LLMs to outline logical steps and execute complex arithmetic computations using integrated tools like NumPy, SciPy, and LEAN4.
- Building and scaling a dedicated "tools dataset" with advanced models to showcase best practices in computational tool integration for enhanced LLM performance.
RF and Optical Communication Course Project
Spring 2023
- Attempted to solve dynamic spectrum access for network utility maximization in multichannel wireless networks.
- Implemented a dynamic spectrum access algorithm that uses multi-user reinforcement learning to obtain the optimal solution for the spectrum access problem with an arbitrarily large-state space constraint.
Real-time Document Localization
Computer Vision and Intelligence Club, Center for Innovation | Fall 2023
- Reproduced the paper LDRNet: Enabling Real-time Document Localization on Mobile Devices by Han Wu et. al.
- Employed Model Pruning and quantization to lower the model footprint by 51% while reducing accuracy by only 1.4%.
Effects of Attention Mechanism on Sequence Learning Models
Fundamental of Deep Learning Project | Spring 2023
- Implemented seq-to-seq model for transliteration using RNN, LSTM, GRU as backbones.
- Implemented Bahdanau Attention in decoder network and compared to it vanilla seq-to-seq models by assessing loss and accuracy measures and visualizing attention maps for qualitative validation.
Implementation of neural network from scratch
Fundamental of Deep Learning Project | Spring 2023
- Used numpy to implement a fully vectorized neural network package used to train a classifier model on the MNIST dataset and Fashion MNIST dataset with wandb support including hyperparameter search using Bayesian optimization.
- Implemented a cross-entropy and mean-squared error with all the major activation functions and Linear layers.
Feedback Prize - Evaluating Student Writing
Kaggle Competition | Fall 2022
- Implemented a BERT based model for segmenting texts and classifying argumentative and rhetorical elements in essays.
- Created a public notebook to help other people understand the problem and my approach, Awarded a Bronze Medal.
Skills
Languages
Technologies
Libraries
Scholastic Achievements
2023
Represented IIT Madras at the Inter IIT Tech Meet and won bronze medal in the computer vision event.
2020
Achieved an All India Rank 657 in JEE(Advanced) 2020 out of 150,000 shortlisted candidates.
2020
Achieved an All India Rank 1779 in JEE(Mains) 2020 out of 1.5 million candidates.
2019
Among the top 10% finalists across the Goa region of the National Graduate Physics Examination.
Leadership Experience
Strategist, Computer Vision and Intelligence Club
Center for Innovation, IIT Madras
April 2022-March 2023
- Spearheaded a team of 40 undergraduate students to make a meaningful impact on real-world issues.
- Engaged with leading businesses, startups, non-governmental organizations (NGOs), and professors during the tenure.
- Organized a range of technical workshops covering fundamental concepts of Computer Vision and Artificial Intelligence.
Event Lead, Generative Model Workshop
Shaastra, Technical festival of IIT Madras
March 2023-Feb 2023
- Mentored the team responsible for conducting workshops on computer vision related topics with over 100+ registrations.
- Ideated and set the content for the session on AutoEncoders, GANs, Cycle GANs, DC GANs and Game theory.