May 2022 – Present · 3 yrs 10 mos · Remote

• Product Category Prediction for Automated Product Mapping
• Designed a retriever-reranker pipeline to classify products into one of 100K+ categories using metadata like titles and specifications
• Fine-tuned a decoder-based retriever; aggregated results using Mean Reciprocal Rank to shortlist candidates across multiple retrievers
• Developed few-shot prompts for a local LLM reranker with single-step decoding; optimized latency using prefix-caching and vLLM
• Achieved 13% accuracy gain, reduced grossly wrong predictions by 66%, and maintained 87%+ high-confidence coverage
• Deployed to production on IndiaMART
• Attribute-Value Extraction from Product Titles and User Queries (Research accepted at KDD 2025 ADS Track)
• Generated weakly supervised training data with incomplete labeling from product specifications
• Designed a novel two-stage system that employs a marker-augmented generative model to identify potential attributes, followed by a cross-encoder based token classification model that determines the associated values for each attribute
• Regenerated training data to expand attribute-value annotations and trained a standard NER classifier on the enriched dataset
• Improved recall by 20% while maintaining 90% precision; used for dynamic feature highlighting to enhance search experience
• Style-Controllable English-to-Hindi Translator
• Extended an encoder-decoder model with style tokens to control translation style; fine-tuned on in-house parallel corpora
• Obtained English translations for scraped style-specific monolingual data using Google Translate API to augment the training data
• Developed a style classifier to sample representative training examples, leading to improved in-style words usage in translations
• Additional Projects
• Dockerized a product search system with large disk-based ANN indexes, deployed using NVIDIA Triton Inference Server
• Developed a shopping assistant to enhance user experience via conversational product recommendations & discovery

Jan 2022 – May 2022 · 4 mos · Remote

• Product Category Prediction for Product Search
• Built a transformer-based classifierto predict the most relevant product category from 100K+ labels using bootstrapped search logs
• Designed heuristics to enhance atomic label representations and sampled data to improve category distribution and coverage
• Achieved similar accuracy (~88%) as the previous seq2seq model while significantly reducing average response time (3x faster) and completely eliminating timeouts; the model was integrated with IndiaMART’s search system and was deployed in production