Advait Save

• Innovated on multi-modal embedding generation for personalized recommendations with improvement of 12% matches online.

• Improved recommendation by +13% clicks by training transformer and graph neural networks, including large scale global deployment.

• Enhanced global ad response prediction by leveraging real-time customer activity and attention-weighted deep interest networks, achieving a 5% increase in AUC and generating +$10MM yearly revenue.

• Led experiments on ad response improvement, incorporating target & context aware neural modules, feature enrichment modules, representation learning, enriched id embeddings resulting in +4% Conversions and +$5MM annual revenue.

• Designed and deployed a contextual ad relevance model using a two-tower neural network, improving ad relevance by +37% and reducing annual inference costs by $2MM utilizing encoder-decoder compression layer.

• Implemented intent detection and Question Answering models in conversational AI bot to answer user queries at scale, improving answer relevance (+30%) and conversions (+5%) for mix-lingual models across India's largest ecommerce audience.

• Trained a domain fine-tuned semantic similarity transformer LLM combined with Retrieval Augmentation Generation from user generated reviews, enabling +40% question coverage and +3.5% revenue.

• Deployed a fine-tuned sentence-BERT to enable multi-modal answers using product videos and images to frequent questions.

• Enabled question-answering across all verticals and languages using multi-lingual machine translation and active learning.