Project Description : Sports AI with NFL, Tennis, Soccer etc
Roles and Responsibility : Implementation of object tracking methods for sports tracking in NFL. Optimizing tracking algorithms to fit sports motion modelling. Strong theoretical understanding of DeepSort algorithm which enabled parameter learning such as mean, co-variance, drift factor, motion matrix, etc. of multi variate gaussian distribution, so that the DeepSort adapts to sports such as NFL, Tennis, Soccer etc with different motion dynamic. MOTA, MOTP metric improvements of tracking algorithms by transfer leaning.
Custom detection of objects for Sport AI (using Yolo v3) using Transfer Learning and further improving it by finding optimized weights by max MAP and IOU scores
Identify weaknesses in trained models using XAI (Explainable Artificial Intelligence) methods such as LIME, Grad-CAM and Influence Function and, improve the trained model
Sports action recognition in NFL using 3D ResNets and able to improve the performance by finetuning it with sports data
Setting-up and implementation of codes in docker and Nvidia-docker environment which matches any target machine environment on which our product is to be deployed
Ceres Solver for converting any 3D transformation problem to non-linear least squares optimization problem
Project Description : Sentiment analysis for e-commerce website based on customers feeedback and deplaoy into cloud with REST API enabled
Roles and Responsibility : Sentiment analysis data collection, data cleaning, data tokenization, trained LSTM model on sentiment data
Wrap python code with flask code, fit the code into docker image, push docker image to Google Cloud Platform (GCP) using GCP APIs. The service was accessible real-time with REST API calls
Handling ML tasks and producing models with fewer data using Transfer Learning Techniques, worked on developing models with NLTK Toolkit, Python Tokenizer and Residual Networks
Published our Node-RED sentiment engine module [link](https://flows.nodered.org/node/neutrinos-sentiment-engine)
Junior research fellow at Indian Institute of Technology Ropar
Tech Stack : Python 2.7.12, Keras, Tensorflow
Project Description : Department of Science & Technology (DST) start-up young scientist project entitled "Activity Learning in Smart Environments"
Roles and Responsibility : Modeling Sequence of Sensor Triggerings in a SmartHome to a Human Activity using Hybrid LSTM+CNN Models with Attention Mechanism. Also, Trigger an Event (such as turning AC on) Based on the Predicted Activity by the Model
Used LSTM Gate Initialization Methods, LSTM Sequence Length Optimization Methods, GridSearch of Hyper-Parameters, and K-Means Cross Validation Techniques to Improve the Performance of the Model
Designed an Encoder-Decoder Type LSTM model with Attention Mechanism to Further Improve the Performance
Experiment with Active Learning Techniques to Build a Strategic Decision Method to Choose Input Data Samples for Training
Summer school in Machine Learning at IIIT Hyderabad
Tech Stack : Python, Keras, Tensorflow
Project Description : Learned about topics such as Optimization, Sequence to sequence learning, GANs, VAE and other generative models, Deep RL and their Applications
Roles and Responsibility : Courses taken by industry and academic experts such as Dr. Tejas Kulkarni (Google Deep Mind), Dr. Ravindran Balaraman (IIT Madras), Dr. Vineeth Balasubramanian (IIT Hyerabad), Dr. Kaushik Mitra (IIT Madras), Dr. Vinay Namboodiri (IIT Kanpur), Dr. C. V. Jawahar (IIIT Hyderabad) and Ankush Gupta (Univ. Oxford).
Assiant Project Engineer at Indian Institute of Technology Guwahati
Roles and Responsibility : Collected Handwritten Data of Characters and Digits using Lightpen on a Tablet PC and Saved in XML Format
Modeled the data using Hidden Markov Model (HMM) based HTK Toolkit
Integrated GUI supplied by CDAC Pune to HMM Model using a DLL file build with Visual Studio (using C). Thus, Integrated Whole System
Visited CDAC Pune for Deployment of the Assamese Handwritten Character Recognition Engine with the Mother Engine
Teaching Assistant at Natioanl Institute of Technology Meghalaya
Tech Stack : C, C++
Work Description : Taken Course and Lab Classes, Designed Lab Question Papers for BTech second year students of NIT Meghalaya for the courses CS 101: Computer Programming and CS 207: Principles of Programming Languages
Indian Institute of Technology Kharagpur
Tech Stack : MATLAB
Project Description : Vehicle Detection and Counting in Traffic Surveillance Videos
Roles and Responsibility : Used Bounded Box and Image Moment-Based Approaches for Detection of Vehicles