Sagar Kaushik

• Trained and integrated a CBOW-inspired recommendation model into the service, achieving 2x the offline recall compared to the popularity heuristic. • Saved nearly $15,000 annually by fine-tuning the model daily and eliminating the indexing API before inference. • Created click-based and inverse propensity score weighted popularity models to rank the rows on pages for cold-start users. • Conducted A/B testing for thousands of users; the weighted model showed 7% better engagement and ranking metrics. • Developed a two-tower model with pairwise learning-to-rank for ranking rows on a page using embeddings of user watch history and rows with their metadata, and achieving 0.72 NDCG. • A/B tested the metadata-enriched row-ranking model, resulting in a 12% increase in successful sessions. • Devised a GPT2 based LLM trained on user watch history data to predict the next watched program, and used that for ranking programs with 0.8 recall @ 10.

• Built a convolutional recurrent neural network based noisy speech classifier which had 73.5% test accuracy. • Trained a transformer model to convert graphemes to phonemes with 6.8% PER, followed by a Text-to-Speech model on the obtained phonemes, and then used a pre-trained HifiGAN vocoder to generate audios.

• Implemented Federated Learning with Differential Privacy to train neural networks for predicting user demographics, obtaining 63% accuracy for age and 76% for gender. • Computed the cost of federated learning and simulated training on various data distribution scenarios on AWS.

• Built a neural network to predict water flow rates for different combinations of reservoir tank valve states with 67% R2.