Multilingual Profanity Detection System• Built a binary classifier for measuring profanity in the Wynk music song comment section• Used a pre-trained DistilBERT model to embed multilingual sentences and fine-tuned the model with additional layers• Achieved F1-Score of 0.92 and Recall of 86%Similar Songs Recommendation System• Trained a Node2Vec model to recommend similar songs for the Wynk Music users• Created an item graph based on user consumption behavior and employed random walks to generate sequences of nodes within the graph using Neo4j• Enriched node representations with songs side features like artist, year, language, and MFCC information• Built ANNOY index on the node embeddings and leveraged these for recommending similar songs• Achieved a metric of 8 Songs Per Session (SPS)

Hyper-Personalized Playlists Recommendation• Created a scalable pipeline to generate hyper-personalized playlists on Wynk Music app for 8 Mn+ daily active users• Trained a collaborative filtering model (ALS) for candidate generation and implemented ANNOY to obtain the top 1000 closest songs for each user• Optimized item embeddings by reducing dimensionality using UMAP algorithm and applied HDBSCAN clustering to craft theme-based personalized mixes tailored to individual users• Achieved metric of 10 Song Per Session (SPS), a Click-Through Rate (CTR) of ~ 17%, and stream contribution of 20%Songs Mood Prediction using Audio Spectrograms• Trained CNN-based classifiers for predicting the mood of the songs• Designed and implemented a comprehensive end-to-end workflow for ETL processes, feature engineering, model training, mood predictions, and dynamic updates to the mood vocabulary within the song catalog• Extracted Mel-scale spectrograms from raw mp3 using Librosa and used these as input to the mood classification models• Leveraged the mood tags obtained from the model to generate mood-based hyper-personalized playlists for users contributing to 3% of the total streams on the Wynk Music appCatalog indexing by popular and trending song prediction:• Developed a comprehensive popularity metric for Songs, Playlists, albums, and artists, considering factors such as streams, downloads, followers, likes, searches, and Hello Tunes data• Implemented an Exponential Moving Average concept to identify trending songs• Applied these popularity and trending metrics to enhance the Wynk music app search functionality, creating popular and trending item lists, and optimizing the ranking of songs in personalized playlists

Bundle Churn Prediction • Built Gradient Boosting Tree classifier to predict potential churn among Airtel’s 100M+ bundle plan customers• Engineered 20+ features based on customer demographics, recharge patterns, data usage, voice usage, and ARPU• Achieved a Recall of 54% and Precision of 70%

Selected through the Mind Over Matter Competition for Internship program INSPIRE – 2017Synthesis of nanomaterial Graphene Oxide using an Industrial waste: Proposed a novel technique for converting coal tar sludge(waste) to an industrial consumable input.Achieved 70% efficiency (~0.7g GO per 1g carbon particle), reduced its cost of production by ~20%.