Kavindra Sharma

Passionate about leveraging cutting-edge technology. I currently spearhead the development of Recommendation Systems for Jio Media Platforms, leading a talented team in this endeavour. My role extends beyond recommendations to include pioneering audio-based features for media content, such as mood understanding and instrument detection, alongside exploring the potential of Large Language Models (LLMs) for the media business.

With an M.Tech in Signal Processing, specialising in audio and speech processing from IIT Delhi, and over 7 years of industry experience in data science, I bring a wealth of knowledge. My expertise spans a range of areas, including Recommendation Systems, Predictive Analysis, Natural Language Processing (NLP), and Computer Vision. Throughout my career, I've had the privilege of leading teams and orchestrating the end-to-end development of systems on a large scale.

Senior Manager Data Science at Jio Platform Ltd. (2018-12 – Present)

• Leads a team of 8-10 in developing and delivering personalised recommendations for JioTV, JioTV+, and JioCinema, handling up to 50 million users for JioCinema and 10 million for JioTV+, with a current aim focused on serving 100 million JioTV users.
• Over 5 years of experience with recommendation systems, managing platform-specific features for live TV (JioTV), OTT content (JioCinema), and set-top box integration (JioTV+).
• Developed various types of recommendations, including user-specific (YMAL, Genre, Language), content-based (similar content), and location-based (popular and trending) for various content types such as Movies, TV Shows, Short Videos, and Live Channels.
• Conducted in-depth research on the functioning of Large Language Models (LLMs), including understanding various parameters and the concept of quantization in LLMs.
• Tuned prompts for diverse use cases, such as enhancing entertainment content metadata and extracting entities and relationships from text associated with content metadata.
• Applied few-shot learning techniques using open-source LLM models like Mistral and Phi-3 to improve classification tasks related to recommendations.
• Deployed Mistral and Phi-3 models as APIs for team-wide use, facilitating broader access and integration into various applications.
• Serves as a comprehensive database for deep understanding of user behaviour, including activity, content preferences, attributes (e.g., language, genre, actors), and usage habits.
• Captures language preferences for individual and regional alignment.
• Supports generating user features, calculating affinities, delivering personalised recommendations, and user segmentation for tailored experiences.
• Enhanced metadata by integrating data from open-source and in-house sources, including deduplication, mapping, and filtering.
• Utilised audio, visual, and text-based models for comprehensive content analysis:
• Extracted keywords and embeddings from descriptions, plots, and subtitles.
• Applied audio models for instrument identification, voice separation, and mood detection.
• Trained various models for movies, shows, music such as
• ALS (Alternating Least Squares): models for movies and TV shows to obtain latent vectors of users and items.
• Bayesian Personalised Ranking: Implemented BPR methods to provide personalised movie recommendations to users.
• Calculated implicit ratings using a percentile-based threshold based on user watch behaviour.
• Utilised metrics such as NDCG, MAP, Hit-Rate, and P@N for offline model evaluation.
• Trained autoencoders on features from item profiles, like genre, actors, director, time of day (TOD), day of week (DOW), and location.
• Used item latent features to identify similar items.
• Applied BERT for generating text-based embeddings as input features from description and tag field of content metadata.
• Performed user segmentation for targeted recommendations using clustering techniques.
• Used various features, including latent vectors from ALS, user profile attributes, content type, interaction type, language, genre, time of day (TOD), day of week (DOW), location, kids' preferences, vendor interactions, and more.
• Applied K-means with k-means++ and evaluated segment quality with silhouette, Calinski-Harabasz, and Davies-Bouldin scores.
• Used KL Divergence to automate segment selection.
• Segments were utilised for Home Page Personalization and behaviour identification.
• Personalised category positioning for user segments by analysing segment affinities (e.g., genre, language, starcast) and category affinities.
• Combined segment behaviour, category behaviour, domain knowledge, and platform behaviour to rank categories for each segment.
• Implementing cross-domain recommendation, connecting news articles with top movies and vice versa, using a knowledge graph.
• Utilising ArangoDB to store and manage the heterogeneous knowledge graph data, which encompasses various types of nodes and edges, including actors, genres, production houses, and news article categories.
• Leveraging DGL (Deep Graph Library) and Stellargraph for training a variety of graph machine learning algorithms on the knowledge graph.
• Experimenting with two prominent algorithms, namely metapath2vec and GraphConv, to explore their effectiveness in this context.
• Split songs into 5-sec segments and use features like Mel Spectrogram, MFCC, Zero Crossing Rate, RMS Energy, Temp, and Pitch. Trained an RCNN model to predict 8 moods based on Russell's V-A Space.

Data Scientist at Hotstar (2016-07 – 2018-12)

• Optimised ad relevance by predicting age and gender for over 5 million users daily.
• Addressed a supervised classification challenge with 8 classes (4 age groups x 2 genders).
• Employed diverse features (numeric, LSI embeddings, clusters) and techniques (PCA, Chi-square, SMOTE) to handle class imbalance and fine-tune models.
• we used the PR curve for getting the correct threshold for a particular class
• MLP outperformed other models such as Linear Regression, Decision Trees.
• Managed end-to-end system performance with robust logging and error reporting.
• Broadened the user base for ad targeting by identifying users similar to seed users.
• Employed various models to score and predict target class likelihood, using lift scores for evaluation.
• Managed the project from conception to production with an automated Python and shell script workflow.
• Done POC for Skip intro and Skip recap features using R-CNN for Hindi TV Series.

M.Tech in Signal Processing – Indian Institute of Technology Delhi, Delhi (2016)

Bachelor of Engineering in Electronics – Madhav Institute of Tech. & Science Gwalior M.P. (2013)

M. P. Board of Secondary Education – Munnalal Agrawal H.S. School, Bhind M.P. (2008)