Devmallya Karar

Experience in building and deploying production-grade, statistically advanced Machine Learning, Deep Learning, Computer Vision and NLP-based state-of-the-art Models for quantitative and qualitative analysis in AWS and GCP. And used GIT for version control of the code in the sphere of product development and services.

Employing state-of-the-art transformer-based Deep Learning models and
Layoutparser OCR model to extract the Named Entity from the insurance
document and stamps for a US-based clients CNA / Hardy and TMK.
Developed a question / answering system using transformer-based model
TAPAS to check the customer's behaviour on the basis of their insurance
usage. Also used LayoutLMv2 model to detect the text as question &
answer on the unstructured data.
With the help of Amazon Textract extracted the text and tables from the
unstructured data. Also used MRCNN object detection model to detect
financial charts from the documents and extracted the financial data.
Managing a team of 12 Machine Learning Engineers and Data Scientists
for this insurance project and engaging with the stakeholders and able to
meet all of their criteria.

Utilized state-of-the-art transformer based DistilBERT NER model to
extract the medical entity from patient insurance online claim forms for
US-based insurers using Spacy and Pytorch resulting in an 88% of
accuracy rate.
Collaborated with a group of MLE and Solution Architects to build an OCR
model using Amazon Textract for the offline scanned claim form and also
utilized MMDetection and Cascade TabNet to detect the tables from the
scanned hospital bills and deployed in GCP using VertexAI.
Pre-processing of image data, applying active learning, built knowledge
graph and fine-tuning of pre-trained models using Tensorflow, based on
time computation and evaluation metrics. By engaging with the
stakeholders, was able to meet all of their criteria.

Utilized Tensorflow and Keras, built an advanced statistically effective
sequential Deep Learning, and time series based prediction model to
estimate retail sales for Walmart, Target, and Pepsico, with a 94 %
accuracy on their quarterly multivariate numerical data and deployed the
model in the Unify BI platform using Flask API.
Built a hand gesture system for a German automobile brand using
Tensorflow Lite and a Deep Learning-based CNN and state-of-the-art
RESNET and MobileNet model which gives 92 % accuracy rate.
Source control of the codebase and build scalable and reproducible
deployment solutions by creating CI-CD end-to-end pipelines.

For an E-Commerce website of HP Inc, had developed a Machine
Learning-based Collaborative Recommendation System with a 94 %
accuracy.
Utilized Django and Heroku to deploy the Machine Learning model, as
well as GCP and VertexAI, and used Postman to check the model's
performance.

MSc in Machine Learning and Artificial Intelligence from Liverpool John Moore's University.