Yash Sirvi

I am Yash Sirvi, a final-year undergraduate in the Department of Computer Science and Engineering at the Indian Institute of Technology Kharagpur. I worked with Prof. Partha Pratim Chakrabarti on scalable learning-based solutions for NP-complete problems, focusing on the Travelling Salesman Problem. Under the guidance of Prof. Debashish Chakravarty, I developed game-theoretic planners for autonomous racing systems. I am also doing my Bachelor's Thesis under the supervision of Prof. Aritra Hazra to improve MARL techniques, addressing scalability and stability in multi-agent settings.

I have also interned at Quadeye Securities LLP as a Systems Engineer. There, I worked on optimizing the compiler pipeline for efficient regression test selection, designing algorithms to detect test and compilation failures early.

Email  /  CV  /  Google Scholar  /  Github

Research

I'm interested in Algorithmic Optimization, Multi-Agent Reinforcement Learning (MARL), Planning, and Robotics, with an emphasis on scalability, efficiency, and real-world applications. I aim to develop innovative algorithms to tackle challenges in autonomous systems and multi-agent learning.

Publications

Decoupling Vertical Federated Learning using Local Self-Supervision
Avi Amalanshu, Yash Sirvi, et al. Presented at NeurIPS 2024 Workshop: Self-Supervised Learning, focusing on fault-tolerant federated learning on vertically partitioned data.

Entity Augmentation for Efficient Classification of Vertically Partitioned Data
Yash Sirvi, et al. Published at IJCAI 2024 GLOW Workshop, introducing novel techniques to handle limited data overlap in vertical federated learning.

DiffClone: Enhanced Behaviour Cloning in Robotics
Sabariswaran Mani, Yash Sirvi, et al. A technical report for NeurIPS 2023 TOTO Benchmark Challenge
Proposing a diffusion-driven approach for policy learning. GitHub Code

Projects

Learning-Based Solutions for the Travelling Salesman Problem
Developed scalable algorithms combining heuristics and graph sparsification to achieve 40x speedup and near-optimal results on large problem instances.
Project guide: Prof. Partha Pratim Chakrabarti (Professor, IITKGP)
With: Maitreyi Swaroop (PhD, MLD CMU) and Sourodeep Datta (CS Undergrad, IITKGP)

Novel Techniques for Scalable Multi-Agent Reinforcement Learning (MARL)
Enhanced SAPG methods for multi-agent environments, integrating centralized training with decentralized execution for scalable solutions.
Project guide: Prof. Aritra Hazra (Assistant Professor, IITKGP)
With: Ananye Agarwal (Phd, MLD CMU).

The Chandrayaan Moon Mapping Challenge – ISRO
Designed deep learning models for lunar surface image super-resolution, enhancing resolution by 16x with cutting-edge techniques. GitHub Code