Vijay Saravana

Performed research under Professor Ling Liu in 3D object detection leveraging Intel NCS2 to perform LiDAR based 3D object detection using Pointpillars by processing video/PCL frames in parallel – achieved mAP of 58 with 2x inference speedup with 2 NCS sticks.

Implemented a MapReduce framework calculating word count from very large text files using gRPC in C++. Performed file sharding and shards were assigned to workers by master for processing. Optimal efficiency was achieved using async completion queues.

Implemented the PAXOS consensus algorithm and built a replicated state machine - demonstrated with Paxos Client and Server. A set of replicas process all client requests in the same order, using PAXOS to agree on the order.

A benchmarking study of NoSQL databases with distributed NewSQL CockroachDB using YCSB workloads by setting up high availability clusters on Microsoft Azure. Implemented automation scripts for performing throughput, latency tests and visualization.

Designed a distributed online store which processes multiple client requests (async) and retrieves bids from multiple vendors (async) using completion queues in C++. This was parallelized with native implementation of thread pool using thread-safe queues.

Implemented various barrier synchronization algorithms in C such as Sense reversing Centralized barrier, MCS Tree barrier, Dissemination barrier using OpenMP and MPI for multithreaded and multiprocess barrier performance evaluations. The project was run in the GaTech PACE cluster and performance was evaluated by scaling the number of processes/threads.

Jinglecraft is a music Genre and Mood classification ML Model. We solved the classification problem using various classifiers such as Random Forest, SVM, Regression and NN - performed a comparitive study among all the classifiers. For Mood classification, we used LastFM dataset and used various classification models (mentioned above) to solve the problem. We also used unsupervised learning techniques such as KMeans to perform clustering of the datapoints.

Pomodoro Time tracker RESTful backend using Java spring boot, Maven, MySQL with CI/CD pipeline implementation using Jenkins. The backend service was used by multiple frontend teams integrating it with mobile applications and websites.

The primary objective of our project is to predict the probability of the occurrence of crime and safety score for a given area at a given time by analyzing historic crime data obtained through crowdsourcing.

Designed a vCPU scheduler for pinning vCPUs to pCPUs based on pCPU usage and a memory coordinator for vCPUs based on thresholds using balloon driver with LibVirt APIs in C.

Built a customisable ERM application using Java RESTful backend, MySQL, JQuery with Redis cache optimization and Kairos for notifications which bagged the runners-up place in Platforms-BU Hackathon.

Android application used for conducting examinations using the smartphone by Network Service discovery to connect the devices of test takers to the examination host mobile device. WiFi Direct NSD removes the dependency on an ISP and AES-128 encryption is used for security.

Detecting `self-deprecating` sarcasm from tweets using the proposed CNN-BiLSTM-DNN network. The proposed concern level intensity classification (CLIC) tree is used to rank users by decreasing level of concern based on tweet history and the details were forwarded to the suicide prevention helpline of India.

Feature ensemble based Sarcasm Detection with features such as emoticons, interjections, internet-slang and lexical featureswith original text corpus was published in ‘Springer AISC 2020.`