Asmita Mukherjee

Development of multimodal retrieval augmented generation pipeline. Ingesting of structured and unstructred data sources(ex: pdf,excel,sql databases,word). Novel iterative approach of multidocument retrieval using gzip algorithm. Significant improvement over the existing OpenSearch solution was achieved. Deployment of Flask Backend maintaining OpenAPI standards with Docker on AWS EC2 instance. Training of BERT based model to tag incident reports to different categories of problmem, cause and classification. Major challenge of handling of multilingual and imbalanced data. Accuracy of 40% was achieved. Development of End to End pipeline to train and evaluate LLAMA model on company data: Synthetic data generation. Custom Tokenization. Pretraining and instruction tuning. Optimizing training pipeline given the hardware constraints. Automated evaluation platform.

Automation tool for monitoring SIP logs generated by skype client to detect and handle potential network or driver issues. Maintenance and support of video conferencing and video streaming infrastructure within Bosch worldwide network.

Project Title: “Web Portal for automating the process of offer letter generation and approval”.

Relevant Coursework: Deep Learning, Machine Learning, Statistical Machine Learning, and Natural Language Processing.