I'm a graduate student pursuing my Masters in Computer Science from New York University.
I see computer science as an equalizer that gives freedom and flexibility to drive real change in the lives of others.“We can only see a short distance ahead, but we can see plenty there that needs to be done.” - by Alan Turing
- Experimenting with different statistical and machine learning techniques to develop a trading strategy.
- Improved the latency of ATDL (Algorithmic Trading Definition Language) strategy tags enrichment by 64% - by optimizing the code, caching, and off-loading the critical trading path (C++, Python).
- Developed key features such as Auto-Settlement, Disbursement, Interest Calculation Engine for the Financial Supply Chain module (Java, Spring).
- Managed and led a 5-member team towards the production support of Payments module.
- Developed a custom machine learning based fraud detection model to predict the probability of on online payment transaction being fraudulent – experimented with Decision Trees, Random Forest, Auto-Encoders etc.
- Developed features such as Debit and Credit consolidation for the Payments module in FinnAxia (Java).
- Promoted to Senior Software Engineer in one year with an excellent feedback.
- Handled 7 million rows of CIBIL (Credit Information Bureau, India) data.
- Determined 15 locations in India with the most potential for a new branch of the bank – using ML algorithms, Microsoft Excel tools and Python.
- Developed contact identity pop-up window feature for the Rexprop CRM Android Application
Explored two different methods for generating images using GANs with only four base images:
1) Meta-training a DCGAN network with Reptile algo.
2) Transfer Learning and data augmentation with DCGAN, cDCGAN, and InfoGAN.
Finally, using evaluation metrics like Mean Squared Error (MSE), Structural Similarity Index(SSIM), and BRISQUE score, we provide the results on which GAN training technique worked best in the context of few-shot image generation.
Technologies: PyTorch, Python, 2 NVIDIA V100 GPU, 16 CPUs
A SaaS (Software as a Service) product which uses computer vision for medical diagnosis. Key technical details:
1) Login functionality: Flask-Login
2) Payment functionality: Stripe API
3) Document storage: AWS S3
4) Computer Vision API: AWS API Gateway and AWS Lambda
5) Database: PostgreSQL
6) Flask app hosted on: Heroku
The competition was to analyze images of galaxies (61,578) to determine the probability that it belongs to particular class.
We tried several different Deep Learning architectures - with different number of layers, optimizers and activation functions.
Our final model was an ensemble of ResNet50 + Xception model (Adam with decay), which helped us climb to rank one on the leaderboard.
Technologies: Keras, Tensorflow, Python, 1 NVIDIA V100 GPU, 16 CPUs
A project demonstrating the concept of client side artifical neural networks. Steps involved were converting Keras model (MobileNet) to TFJS model (tensorflow.js), serving models with Node.js, training and transfer learning in the browser, and finally deploying it to cloud (Heroku).
The model predicts the top 5 prediction probabilities for the image uploaded (out of 1000 trained classes)
Technologies Used: Tensorflow.js, Node.js, Keras, Express, Heroku
An android app which classifies the FRIENDS sitcom cast
An android app which classifies the camera display image to the FRIENDS sitcom cast in real time. The 600 images scraped from google cloud were trained on top of MobileNet model (Transfer learning).
Technologies Used: TensorFlow Lite, Keras, Python, Java, Android Studio
An app which plots the probability of an image uploaded to be a daisy, rose, or a sunflower.
Technologies Used: Keras, Flask, Bokeh, Heroku
Implemented the following operations: Forward Propagation, Cost Function, Backward Propagation, Parameter updates, Train, Predict, Affine-forward, Affine-backward, Activation forward and Activation backward in python using just the matrix operation library (numpy). Did not use any other python libraries/modules. (only used Pytorch to load the data)
Technologies Used: Python (numpy), Pytorch - for data loading
Research paper presented and published at the International Conference on Machine Vision and Information Technology (CMVIT 2017) organized in Singapore.
Technologies Used: Python/R, Neural Networks, Gradient Boosting, SVM
Research Paper presented and published at the 15th International Conference on Applied Computer and Applied Computational Science (ACACOS 2016) organized in Prague, Czech Republic.
Technologies Used: R, Neural Networks, Random Forest, Ensemble Techniques
I took some leadership oppurtunities in the organization which include:
1. Organizing Committee member for January Recruitent in 2014.
2. Organizing Committee President for February Local Congress in 2014.
3. Organizing Committee President for Youth Carnival, Annual AIESEC Youth Fest in 2014.
Have won various inter-college and state level competitions in the singles and doubles category.
