- Built a python package based on pytorch, making easy to run state-of-the-art deep reinforcement learning algorithms.
- Implemented algorithms such as DDPG, DDQN, CEM. pytorch-rl works with OpenAI Gym out of the box.
- keras-3dgan | 3D Generative Adversarial Networks
- Implemented a NIPS 2016 paper on 3D Generative Adversarial Networks to generate 3D Objects in keras.
- Experimented with inverted loss, autoencoder pretraining and bias-free versions to overcome the GAN convergence problem.
- Hyperspectral Satellite Image Segmentation
- Worked on hyperspectral (4 and 20 band) high-resolution satellite image segmentation using convolutional encoder-decoder neural networks like SegNet.
- Developed an end to end pipeline achieving a mean accuracy of 82% on Zurich dataset.
- Carrom-RL | Reinforcement Learning
- Developed a reinforcement learning agent to play Carrom in single user and multi player settings.
- Explored and implemented reinforcement learning various techniques as deep Q-learning (DQN) and DDGP.
- Segmentation and Classification of Web UI elements
Guide: Prof. Antti Oulasvirta
- Developed an adaptive system to segment and classify personalized UI elements
- Achieved 91% F1 score in segmenting web UI elements using edge-detection based techniques and other heuristics
- Classified the common web elements with a mean accuracy of 97.4%
- Incremental Query Optimization
CSE, IIT Bombay, Summer 2015
Guide: Prof. S. Sudarshan
- Worked on incremental search algorithm along with branch and bound pruning of the search space to improvise the Depth First Search approach used in the Volcano and Cascades Frameworks
- Implemented Fibonnaci heap for the dynamic execution of the Task depending upon the priority
- Improved efficiency of the cost metric in the PyroJ query optimizer to achieve the best plan for execution
- Image Recognition Using Class Specific Linear Projection
Digital Image Processing, CS 663
Guide: Prof. Ajit Rajwade
- Implemented the algorithm for face detection, insensitive to lighting direction and facial expression
- Image Mosaicing Using Fourier Shift Theorem
EE338 : Digital Signal Processing
Guide: Prof. V.M. Gadre
- Used an algorithm based on Fourier shift theorem for image mosaicing
- The project was ranked among the top 3 in the whole batch.
- Swarachakra Bangla
IDC, IIT Bombay, Summer 2014
Guide: Prof. Aniruddh Joshi
- Improvised the existing Swarachakra code-base to develop a version in Bangla
- Conceptualised the design and layout of the keyboard for a better user experience
- Swarachakra Bangla has attracted over 100,000 downloads on Google Play Store
- ASC Visualization Kit
Software Development For Engineers, AE 425
Guide: Prof. Shankar Balachandran
- Developed a Django application to enable the student community to have a better access to academic data
- Used data scrapping techniques viz. Selenium for collecting data from the web
- Winner of the Facebook hackathon for the best utility project in the institute