Swapnil Kumar

I am a Graduate Software Engineer Trainee at Jaguar Land Rover. I graduated from IIT Bombay in 2020 with a major in Mechanical Engineering and a minor in Computer Science and Engineering. My interest lies in areas of Artificial Intelligence with a focus on Computer Vision, Numerical, and Robotic Simulations. I undertook various projects, courses, and experiences in these fields. I am also fascinated by other fields, including distributed computing, robotics, and optimization. I was a member of IITB Mars Rover Team and also worked on developing an efficient, accurate and versatile Human Activity Recognition system in the Machine Intelligence Program under the supervision of Prof. Asim Tewari.

IITB Mars Rover Team

I was a member of the IITB Mars Rover Team, in Mechanical Subdivision. I primarily worked on design optimization and simulation-based analysis. I spearheaded the design of an ambitious 4 wheeled dependent suspension system for the rover. I also worked on the simulation-based optimization of the component designs on ANSYS. I contributed to the development of Universal Robotic Description Format (URDF) of the Rover and implemented its simulation in Gazebo, to perform dynamic analysis of mechanical subsystems in various environments (worlds).

Research Intern
JFE Steel Corporation

I worked on the project to simulate the steel refining process in a converter. The pioneering model simplifications limited the convective-courant number in high-velocity regions below 1, resulting in solutions with significantly improved computational time and accuracy in line with the experimentally observed values. The effects of key design and functional parameters on the efficiency and time of the process were researched and simulated. Suggested converter and outlet design improvements, effected 60% reduction in process time, and 10% increment in process efficiency.

2-Stage Human Activity Recognition
Prof. Asim Tewari

An efficient, modular, and versatile Human Activity Recognition system, a new 2-Stage approach has been proposed. The approach uses human joint localization to estimate joint angle variations in time, which is used for recognition. The model is trained on KTH datasets, it gives an accuracy of 98.19% which is higher than state-of-the-art models at a less computational cost. The model is also extended to work for multiple agents in a frame and possible applications in the Assembly line safety is explored. For more information, please refer to the report

Multiphase Modeling of mould filling in Epoxy Resin casting process
Prof. Abhilash J. Chandy

We developed a simulation for the process of Resin injection in mould during the casting of an insulation layer on a transformer core. The objective was to study the reason for cracks and cavity formation while casting. Design modification was also suggested to make the process efficient. The work in this project was presented in Paper No. 490 of Fluid Mechanics and Fluid Power (FMFP) conference, 2018

Quaternion-based Model for Human Motion
IE643: Deep Learning: Theory and Practice

The Quaternet model predicts future human poses by studying Qaternions based representation of temporal joint angles. A new underlining seq2seq architecture was proposed and implemented to improve computational efficiency. This architecture reduced the training time by 53.6%, without any significant change in the accuracy. For more information, please refer to the report.

Institute Technical Summer Project
ITC, IIT Bombay

Fabricated a bot named 'Zero-G', that moves around by pressure gradient of the propellers. It's Servos and brushless DC motors are interfaced with the Arduino UNO micro-controller package and get commands from an Android app using the HC-05 Bluetooth Module.