Built the entire pipeline for question to question search. We takes the question and find the similar
question from our dtabase to map it with the corresponding Learning Objective.
Responsibilities:

1. Design entire structure and implemented the pipeline for search functionalities.
2. Took the entire mapped data and build faas model on top the embedding vector to search.
3. Created the api for this project and deploy it for the use of content team.

Built the moderation pipeline using deep learning model which can automatically moderate
the content for the forum. This model resticts the negative content to be posted on the forum.
Responsibilities:

1. Clean the entire data and build the dataset for model training including synthetic data generation.
2. Building model with different set of approach and compare result with different Approaches.
3. Build the entire pipeline starting from getting the data and filter it by using moderation.

Correcting AI (Image Detection Correctness using Representer Function): Explainable AI (XAI), Pytorch, YOLOV3, Representer Function

• Explaining and Correcting a Deep Learning Detection model by changing the input training data with the help of Representer Function Algorithms. We have designed and developed an algorithm on top of YOLO algorithm and Representer function Algorith which can explain the correctness of the detection model. It includes modifying the YOLO V3 loss function and YOLO V3 optimisation function with its implementation.

Responsibilities:

• Proposed the use of Representer function for to increase the Object Detection model accuracy.
• Build different object detection model for the framework.
• Understanding and designing complete YOLOV3 architecture, training, optimisation
• Process and building of different YOLOV3 model for different dataset.
• Research and development of YOLO loss function Implementation for Representer function and different other methods for the project.
• Design and Development of the new algorithms for the framework.

Detecting the existence of brain tumours from MRI scans and Segmentation of the tumour region by using Deep Learning. Used ResNet-101 architecture to train the Deep Learning model which can detect the existence of brain tumour and used U-Net architecture for the Segmentation of tumours.

Classification of sentences into different categories such as positive, negative, and neutral on indian languages using sentiment analysis.

Responsibilities:

1. Built the classifier model to classify the sentences in to different categories.
2. Used Word2vec and Fasttext embedding for indian language and also handled unseen word problem.
3. Used CNN and LSTM to build the model.

Deep Learning Library LRP Generator: Tensorflow, Torch, Keras, LRP, Python

• Built a Deep Learning Framework to generate heatmap (from LRP) independent of any architecture, any framework (currently supports TensorFlow, Pytorch, Keras) which includes implementation of all the deep learning layer from scratch for ex: CNN, FC, ReLU or Pooling layers.

Responsibilities:

• Understanding of Layer-Wise Relevance Propagation algorithms and implemented it in for our product with customisation.
• Architectural, Designing and complete development of the module from scratch.
• Tested it for around 20 different model.