Cloud / Ops:Architected and implemented the splitting of deep learning models into separate service from core code; setup CI/CD pipeline for models using DVC and MLFlow (allowing for versioning, rollbacks and reproducibility for models) Setup Kubernetes cluster with Terraform and deployment of deep learning models with AWS CodePipeline and KubernetesSetup CI/CD for backend using Docker, AWS CloudFormation and AWS CodePipeline adhering to immutable infrastructure best practicesSetup Auto Scaling Groups to allow for rolling updates and high availability with lifecycle hooksSetup HIPPA/GDPR compliant AWS infrastructure, with VPCs, logging etc. Computer Vision:Implemented and integrated pretrained encoders, such as ResNet, DenseNet, and Squeeze-Excitation Nets for our FCN model which improved IoU by 30%-40%Implemented mechanism to allow FCNs to produce image embeddings which are used for active learning to query for similar images in the unannotated datasets that the model is unsure about, which reduces the time spent on annotationImplemented various Fully Convolutional Network architectures for segmentation of bone images from CT scans (2D/3D UNet, DCAN, FCNs with CLSTM) in PyTorchImplemented our deep learning training toolbox, which allows for rapid experimentation over various model architectures and pre/post-processing techniques. Includes features such as automated generation of decoder architectures for a given encoder (for FCNs), semi-automated learning rate finder, all easily configured via a configuration file.

Implemented a CNN to do automated aesthetics prediction of images achieving ~0.7 correlation with human ratersImplemented traditional and deep learning (mainly seq2seq with Attention, and Transformer module) methods for text generationImplemented Python backend server, using RabbitMQ and MongoDB, for clients to interface with text generation neural network, and moved server infrastructure onto EC2.Full-stack app development, using React Native front-end, Node.js as backend and MongoDB for database stack

Optimization of genomic models for predicting complex traits: Used glmnet in R to create and find the optimal hyper-parameters of LASSO, ridge regression, and elastic net through cross validation for the prediction of complex trait for different genetic architectures Meta-analysis of models for predicting complex traits - Compared the predictive power of different models (BLUP, MultiBLUP, and SparSNP) where BLUP and MultiBLUP are models based on ridge regression, and SparSNP is based on LASSO