Hello World

Spearheaded the development of Apache Beam pipelines for building out Retrieval-Augmented Generation (RAG) applications in collaboration with Google Cloud’s Beam Team targeted at increasing user acquisition. Implemented two key pipelines: one for ingesting and indexing text corpora into a vector database, and another for semantic search, enriching user queries with relevant text chunks from the knowledge base. Demonstrated proficiency in MLTransform for preprocessing data, generating embeddings, and enhancing data with Beam's Enrichment transform, contributing to the advancement of semantic search-based applications in the ML community.

Worked on this project as part of Google Summer of Code 2023 Contributor.Worked on Apache Cloudstack (IAAS at the intersection of cloud computing, Virtualization and DevOps) for extending export/import unmanaged instances for KVM Hypervisor. Work involved understanding VMWare specifics and then adding functionality in the management server and KVM Plugin to facilitate API+UI support for listing, unmanaging and importing external VMs.

The Q-A Bot project aims to create an advanced Question-Answering (QA) system using Retrieval-Augmented Generation (RAG) techniques. This system is designed to answer user queries accurately by leveraging external knowledge bases.

Recommend safest and fastest route

Netflix is one of the most popular media and video streaming platforms and this Visualization is helpful elpful in Data Driven Decision making in Entertainment Industry and understanding viewer preferences and behaviors to improve content recommendations and enhance user experience.

The BRAINTEASER task centers around examining elements of speech that are not immediately apparent. The input is a question and a list of choices. Notable about the question is how it takes advantage of unconscious assumptions and irrelevant information. The dataset is in English, and provided by the task organizers. For the brainteaser task, there are two subtasks: sentence puzzle and word puzzle. Sentence puzzles focus on unstated, typical, and wrong assumptions about the question/answer, while word puzzles defy common sense since they hinge on a word's actual alphabetical makeup rather than its actual denotation. We chose to focus exclusively on the sentence puzzle. We made use of BERT, XLNet, BART,T5, yoso and Roberta models specifically the base HuggingFace implementations of each models.

In this Project I experimented with various NLP based classifiers and used Softmax thresholds to identify one of many toxic attributes in text (toxic, severe_toxic, obscene, threat, insult, identity_hate). I Built various models to benchmark on Vanilla RNN, LSTM, GRU and DistilBERT. Also attempted fine tuning OpenAI APIs based GPT-3 models. Achieved best results of 98.3% using GPT-3 (Ada base) fine tuned model.

Tweets Sentiment Analysis on Abortion Rights

Monitoring the greenhouse gas emissions can be a time and resource consuming task for a lot of the industries. Both periodic measurements and continuous emission monitoring systems can be cost intensive as they would require constant monitoring of the sensors used to measure the gas emissions. Research was published by the Turkish journal of Electrical and Computer Science department to tackle this issue. The research proposes a predictive emission monitoring system (PEM) that predicts the emissions of CO and NOx using calibrated equipment established on site. The researchers set a benchmark for the predictive models built and this project is an attempt to apply statistical methods for the data.