Jul 2024 – Dec 2025 · 1 yr 5 mos · Bengaluru, Karnataka, India

• Search relevance and reranking
• Led the development of transformer based candidate generation models to improve catalog search across head, torso, and tail queries. Built and productionized two-tower models trained on 200M+ query-catalog pairs to power multimodal and vernacular search. Today, majority of the feed you see in Meesho search is powered through the candidate generators built by me
• Launched three major model versions in production:
• v1: Multimodal model using image-text embeddings, driving +1.05% NMV/Visitor platform, +11% torso and +4% tail query conversion, and improving catalog diversity with +1.3% OC uplift.
• v2: Added novel auxiliary tasks for improved vernacular query understanding, boosting tail query conversion by +10.6% and voice search by +14%.
• v3: Developed a shared encoder architecture using VLLM-based catalog summaries, improving semantic alignment and driving +2.88% Orders/Search on long-tail queries.
• Also benchmarked multiple multilingual models for Indic languages, built synthetic training data creation pipelines and deployed LLM-based cleaning pipelines to scale training data quality
• Tech Stack: Pytorch, Pyspark, GCP Vertex AI

Mar 2022 – Jun 2024 · 2 yrs 3 mos · Bengaluru, Karnataka, India

• Architected and developed both projects from scratch
• Product Match: Deployed a large-scale Vision Transformer-based image matching system to identify
• duplicate and visually similar products across 150M+ listings, serving as a foundational service for
• pricing, cataloging and compliance.
• Multimodal Representation Learning: Developed unimodal and multimodal (image-text) embeddings to
• power key downstream systems including search relevance, catalog-catalog recommendations, and
• user-item recommendations. Built baseline models and finetuned 500M+ param open source transformer based models on large scale training datasets
• Tech Stack: Pytorch, Pyspark, AWS Sagemaker

Sep 2020 – Feb 2022 · 1 yr 5 mos

• Worked on some interesting problems:
• GIS:
• Built locality/sublocality polygons from intersection of road and rail networks using open street map data
• Explored various open sourced datasets and used them to build indexes that can help product team with greater insights
• Solved for incorrect customer locations using historical pings from delivery partners
• Address Geocoding: Given an address string map the address to a network path in the supply chain. Used preprocessing techniques like probabilistic splitting of compound tokens, spell correction, bigram merger and trained a classifier to map an address vector to it's right network path. Explored application of information retreival systems to help solve for the same.
• Complaints Prediction: Given an order estimate the likelihood of the order leading to complaint. Used various features associated with an order and captured their interaction to predict orders that will lead to complaint.

Dec 2019 – Sep 2020 · 9 mos

May 2019 – Jun 2019 · 1 mo · Bengaluru Area, India

• Built a knowledge graph of interaction of users with Wysa which is used in the downstream task like building collaborative filtering recommendation engine and learning user behavioural patterns.
• Upgraded the nlp codebase from python2 to python3.

Jan 2017 – Jan 2018 · 1 yr · India

Jan 2016 – Jan 2018 · 2 yrs · India

Jan 2016 – Jan 2018 · 2 yrs · India