Aug 2021 – Dec 2021 · 4 mos

• Was associated with the SRP project: Designing Efficient Hardware Accelerators for Autonomous Driving Vehicles (ADVs). This project is led by Professor Sparsh Mittal from IIT Roorkee. This is in conjunction with my work at IIT Roorkee and involves making neural networks used in ADVs robust to any form of adversarial attacks.

Jun 2021 – Jul 2022 · 1 yr 1 mo

• Researching at the Tech4Good Lab at UCSC under Prof. David Lee.
• Working on using reinforcement learning(RL) to analyze economic situations for apprenticeship learning where apprentices are mentored to work on projects under the guidance of a mentor.
• Leading a team of 4, creating an intricate environment requiring multi-agent RL with mentors and mentees, a planner to supervise the training policy, and generating projects while addressing complex scenarios arising in the real world.
• Implementing actions an agent has to take simulating the real world with deadlines, rewards split between agents working on the same project and agents working on multiple projects simultaneously. Developing the RL algorithms.
• Setup and managed the Pacific Research Platform Kubernetes cluster granted for the project

Apr 2021 – Jun 2023 · 2 yrs 2 mos · Roorkee, Uttarakhand, India

• Working under Prof. Sparsh Mittal at the CANDLE Lab
• Worked on adversarial robustness, proposed novel techniques for defending models against adversarial weight attacks.
• Developed a new adversarial weight attack that can bring down the accuracy of state of the art models trained on ImageNet to 0.2% by changing values of less than 25 out of millions of trainable parameters in a model.
• Developed a targeted attack that brings down accuracy of models on one class while retaining accuracy on other classes.
• Working on a novel adversarial input attack using a region specific approach to bring down the accuracy of an attacked network by attacking just 16-20% of pixels in an image in contrast to methods like FGSM that perturb all pixels
• Developing a GAN based defence approach against multiple adversarial input attacks. The defence strategy maintain an accuracy of 62-67% when an attacked image is fed to the system. Without the defence accuracy drops to a near random guess accuracy of around 1-2%.

Sep 2020 – Dec 2020 · 3 mos

• Worked on 6 sprints - 1. Open Ended 2. Education 3. Game Dev 4. Data, AI and ML 5. Dev Tools 6. Social Good.
• Was one of around 170 students selected for batch 1 of the fellowship out of 20000 applicants.
• Worked on projects such as mobile apps for aiding visually handicapped people, a mentorship system, and much more with a variety of technologies such as Django, React, Flutter, Godot, etc.
• My projects were judged amongst the top 3 in four out of six sprints.

May 2020 – Sep 2020 · 4 mos

• Trained single shot detection models to detect and classify objects and building materials in household environments.
• Due to limited data available, performed several augmentation techniques to prevent overfitting and converted trained models(MobileNetV2 architecture) to tflite models with an object detection accuracy of 91.3%. Built an Android app to leverage Google MLKit for using tflite models and automatically fill Poverty Probability Index(PPI) surveys.
• Wrote the documentation for the entire project.

Nov 2019 – Jan 2020 · 2 mos

May 2019 – Aug 2019 · 3 mos

• Developed the hardware simulation, Badge Magic Android, of a LED name badge, by passing the 2D array into a filter of animation specific algorithm; this enabled people without the hardware to experience the hardware beforehand.
• Worked on the Phimp.me application which is photo editing tool and implemented new features, filters and effects.
• For both of these apps, I automated PlayStore and F-droid deployment process and improved the build time by 5minutes using Fastlane tool, bash scripting, and continuous integration

Nov 2018 – Jan 2020 · 1 yr 2 mos

• I contributed to the Phimp.me and Badge Magic Android projects under FOSSASIA. Along with this, I also contributed to SUSI.AI and Open Event Android.

Nov 2019 – Jan 2020 · 2 mos

Jul 2019 – Aug 2022 · 3 yrs 1 mo

Jul 2018 – Oct 2022 · 4 yrs 3 mos · Kerala, India

• amFOSS is the Free and Open Source Software club of my college. I have been an active member of the community from the time I joined college. I actively take part in all events and also help in organizing events hosted by amFOSS. I help manage our social media campaigns, club projects, setting up pipelines for several subteams, and much more. I mentor my juniors and get them exposed to new technologies and open source as well.