Driftly is an AI-driven social media assistant that helps users write, schedule, and automate social media posts. Users can generate engaging content using AI and schedule posts for optimal engagement times. Additionally, users can provide a set of links, and Driftly will scrape data from them to automatically generate and post content daily or weekly.

Plato is a high-level abstraction over LangChain that makes it easy to build, configure, and deploy AI-powered agents. It provides a structured way to manage:

• LLMs: Seamless integration with multiple language models

• Tools: Simple tool management and integration

• Memory: Efficient conversational memory handling

• Workflow: Simplified orchestration of complex workflows

With just a few lines of code, developers can: • Create multi-agent systems • Integrate external tools • Automate AI-driven tasks

All of this without dealing with the underlying complexity of LangChain.

AI-Powered Visual & Semantic Search enables users to find products effortlessly by leveraging both images and text. Built with OpenAI's CLIP, it understands product images and descriptions in a shared embedding space, ensuring highly relevant search results. Users can upload images or describe products naturally, and the system intelligently retrieves matching items from the store inventory.

Shop The Look allows users to upload a photo of an influencer or celebrity, and our AI system will identify the clothing & accessories worn in the image. Then, the app will find similar products from the store and display them for purchase.

SayNSave is an Android app that lets users record their voice, automatically transcribe it into text, and rewrite it into a beautifully formatted PDF. Users can store, revisit, and listen to past recordings, as well as access their saved PDFs anytime.

A Flutter-based AI-powered news app that curates and summarizes the latest happenings into 2-minute digestible updates. The app intelligently tracks user interactions, such as which news articles they pause on and read, to refine recommendations over time. By leveraging AI-driven content analysis, it ensures that users stay informed with the most relevant and engaging news stories - based on the trends from x, insta and other social platforms as well. The app features AI-generated audio narration of news summaries, allowing users to listen to the latest news on the go. News content is sourced from newsdata.io API to ensure comprehensive and up-to-date coverage.

Developed Voyage, a route-planning and booking system that processes OpenStreetMap geo-data to identify train stations, airports, and bus stations. The system finds the most efficient path between any two locations, optimizing across multiple modes of transportation (trains, flights, buses). Additionally, users can book all required transport options in one seamless transaction.

Developed an LSTM-based deep learning model for AI-generated music composition. The model learns musical patterns and structures from training data and generates coherent, melodious sequences. By leveraging sequence-to-sequence modeling, it produces realistic AI-composed music, demonstrating the potential of deep learning in creative applications.

Trained an EnlightenGAN-based low-light enhancement model to improve visibility in real-life low-light conditions. The model was tested on various real-world datasets to evaluate its effectiveness in enhancing dark images without overexposure or artifacts. The results demonstrated the potential for real-world applications in photography, surveillance, and night vision enhancement.

Developed a product recommendation system utilizing sentiment analysis of user comments and implemented T-SNE with cosine similarity to find ingredient similarities. The project also features a virtual try-on system using BeautyGAN, allowing users to visualize how products would look on them. This project was developed as part of the Myntra Hackerramp competition.

During my internship at Samsung Research, I published research in the International Journal of Speech Technology (SCI journal) focusing on improving ASR text analysis for named entity identification and categorization. The implementation achieved significant improvements in NER models, with Base BERT NER F1 score increasing from 94.73% to 95.37%, and RoBERTa NER from 94.13% to 95.14%.

This prototype aims to demonstrate a driver safety application that ideally should be integrated into vehicles themselves, providing enhanced control over braking and other critical functions. It was selected as one of the top 50 semifinalists in the Google Solution Challenge 2021. The application features drowsiness detection, traffic sign recognition using SSDMobilenet, and an accident detection algorithm that triggers automatic location-based alerts. Performance optimizations have reduced object detection time from 600ms to 96ms per frame, significantly enhancing real-time safety measures.

This project focuses on enhancing Entity Linking by integrating BERT similarity with the existing ABACO method. Entity Linking is the process of mapping textual mentions to real-world entities, and this work aims to improve its accuracy using deep learning techniques. By incorporating BERT similarity alongside BM25 ranking, the project significantly improves entity disambiguation, making the linking process more robust. The approach leverages DBpedia, Wikipedia APIs, and custom graph-based methods to refine entity selection. This project was submitted as part of the CSE2003 Data Structures and Algorithms coursework at the School of Computer Science and Engineering, Nov 2020.