Core objective was to build a system for the video search analytics that can generate different events on the suspicious activity. The application was for law enforcement agencies who can proactively identify suspicious events and have actionable information to support it. These incidents were captured via a high-volume Kafka event and videos were being scrapped from the different social media platform such as Facebook, snapchat, twitter and others. The processing engine was downloading the videos and then it was converting into frames and performing different models which extracted text, objects, face, and other information from these frames. The project included work on the below areas. 1. Setting up Kafka for capturing high volume change events. 2. Scrapping videos from social media platforms (i.e., Facebook, Twitter, Snapchat)3. Architecting the project into different microservices. Each microservice designed to deliver a function independently.○ Frame extraction○ Image quality improvement○ OCR for Text extraction on every 10 frames○ NLP model for Text quality improvement ○ Object detection using Detection on every 10 frames4. Local version of S3 bucket was used (i.e., Min.IO) for image and videos storage.5. MSSQL database is used for storing extracted information.6. Configured MSSQL CDC with Kafka brokers.7. Celery Workers is used for long and extensive tasks

Collection of data from different sources, including but not limited to, Google analytical data, Sales information from local application DB (MariaDB), Scraped data from Travel insurance website of UK, US and Canada. The changes were coming in from Travel insurance sources and sales transactions were captured via the Kafka events and processed using our data cleansing microservices. Our processes were creating intelligent consumable information by co-relating Sales data with Analytics. Below areas were focused.1. Use of apriori algorithm to generate user recommendations in ITIRA.2. Data cleaning of sales data for recommendation system.3. Management all the insurance sales data.4. Collection of the user information from google analytics, AdWords, and other March 2019– May 2019April 2017 – March 2019April 2012 – March 2017(with gaps of MS)marketing tools.5. Correlation of all the sales data with marketing data to figure out the KPI performance.6. Define and implementation of business KPI related to sales, marketing, and site performance.7. Visualization of the information using different sources like sales data, users metrics data on the G1G dashboard.8. Leading and assisting team in selecting the data science stack for development

The project was to reduce noise during the conference calls made via the Confu application. This required research and implementation of noise reduction and quality management techniques. The below areas were focused during this project. 1. The key role is to produce prototype solutions and meaning full research for my team to implement those solutions.2. Build the SIP-based server that connects the VOIP calls from the user's end.3. Done extensive research on RTP, SIP, and other related VOIP protocols.4. Implement the algorithms for Active Echo Canceler (AEC), Active Noise Suppressor (ANS) to enhance the audio quality, and also it will help in an android application where these said algorithms are not supported on some Android devices. (Stack: Python, C++).5. Reduce the residual noise from the voice signal, Deep learning-based RNNoise is implemented at the server-side ( EC2 Amazon ) (Python, Tensor-flow-GPU).6. Encoder/Decoder is implemented in C++ using Opus library to consume less channel bandwidth.

HeatMap: Developed the product Heat Map as a retail analytics product, which infers the realtime sales behavior in a store. the standalone system generates the heatmap and sends the result to Dynamodb. Indigenously built the Stereo Camera for computing the depth map.(Technologies used: OpenCV, Dynamodb, S3, RTSP protocol).FootFall: This feature tells the net customer who visited the mall/store. The system processes the stream and saves counts for the individual pedestrian. Face detection is implemented using tensor-flow with the pre-trained model SSD(single shot multi-box detector) trained on the WIDER FACE dataset. Implement the blur detection on the face to optimize the speed of the programs. Implement the tracking algorithm using Dlib OpenCV for pedestrian tracking and counting (Technologies used: OpenCV, Dlib CT, Caffe Detector, Dynamodb).Real-Time Ad Playing System: Developed the indigenous in which the ad is playing based on gender prediction. The stream is retrieved from the pi’s cam and our train model predicts the gender and triggers the ad on the android application. Watershed Algorithm used for motion detection in the video stream. Server to Pi’s communication via TCP/IP protocols.

Teaching undergraduates below subjects1. Computer science in General2. Signal & Image processing3. Computer Architecture4. Control engineering