CV
Updated January 2025
Basics
Name | Kaustubh Ponkshe |
Label | AI Researcher |
kaustubhponkshe@iitb.ac.in | |
Phone | +91 7888066398 |
Url | https://github.com/KaustubhP11 |
Summary | A dedicated researcher specializing in federated learning, privacy-preserving AI, and efficient model fine-tuning, with experience at prestigious institutions including MIT Media Lab and MBZUAI. |
Education
-
2019.08 - 2024.07 Mumbai, India
Interdisciplinary Dual Degree
Indian Institute of Technology Bombay
Electrical Engineering and Artificial Intelligence
Work
-
2024.11 - Present Abu Dhabi
Researcher
Mohamed bin Zayed University of Artificial Intelligence
Working on LoRA-SB and FedEx-LoRA projects under Prof. Praneeth Vepakomma
- Developed novel initialization strategy for low-rank adaptation
- Achieved 90x parameter reduction while matching performance
- Formulated systematic error residual assignment for federated LoRA
-
2023.05 - 2024.05 Boston, MA
Research Intern
MIT Media Lab, Camera Culture Group
Worked on Power-Learning project focusing on differentially private embeddings
- Established theoretical conditions for private data distribution
- Achieved 5% improvement over DP-SGD while maintaining privacy guarantees
- Developed joint learning framework for privacy-utility optimization
-
2022.05 - 2022.07 Japan
AI Research Intern
AWL Inc. Japan
Worked on Core-AI-Engine optimizations
- Implemented parallel processing for face detection and classification
- Developed three graph-based optimization algorithms
Publications
-
2024.11.01 Initialization using Update Approximation for Extremely Efficient Low-Rank Fine-tuning
Under Review @ ICML 2025
-
2024.06.01 FedEx-LoRA: Exact Aggregation for Federated Fine-Tuning of Foundation Models
FITML Workshop @ NeurIPS 2024
-
2024.04.01 Power Learning: Differentially private embeddings for collaborative learning
Trustworthy ML Workshop @ CIKM 2024
-
2024.03.01 StructFormer: Document Structure-based Masked Attention
Under Review @ TMLR
Awards
- 2024.05.01
Undergraduate Research Award (URA 02)
IIT Bombay
Awarded for one of the best bachelor's thesis in the entire batch
- 2019.01.01
Projects
- 2024.11 - Present
LoRA-SB
- 90x parameter reduction while maintaining performance
- Theoretical guarantees for gradient optimization
- Applied to models up to 9B parameters
- 2023.07 - 2024.07
GUIDEQ
- Novel occlusion-based method for keyword identification
- Improved F1-scores through enhanced information collection
- Applied to healthcare and legal domains
Volunteer
-
2022.07 - 2023.05 Design Engineer
IIT Bombay Racing
- Designed autonomous controller for FSSIM simulation
- Won Software Intelligence competition by Ather Energy
- Implemented dynamic control systems
-
2020.06 - 2021.05 Sports Convener
Institute Sports Council
- Organized Blackcat Championship for varsity teams
- Executed Virtual Run with 1500+ participants
Skills
Machine Learning | |
Federated Learning | |
Privacy-Preserving AI | |
Model Fine-Tuning | |
Deep Learning |
Programming | |
Python | |
TensorFlow | |
PyTorch | |
Simulink |
Mathematics | |
Optimization | |
Linear Algebra | |
Statistics | |
Calculus |
Languages
English | |
Fluent |
Hindi | |
Native |
Marathi | |
Native |