Oct 2024 – Apr 2026 · 1 yr 6 mos · Bengaluru · Hybrid

• Project Lead for Trends Charter: Search "trendy tops for women" on Flipkart App → trends surface as Hashtags (e.g. #CowlNeck, #Draped).
• 1. Trend Identification: 20+ rising trends weekly from Instagram posts of Indian fashion influencers and brands
• 2. Trend to Product Tagging: Built 2-stage Retrieval + Reranking pipeline (CLIP + Qwen-3-VL-Reranker); fine-tuned reranker using LoRA SFT; deployed using vLLM

Jun 2021 – Oct 2024 · 3 yrs 4 mos · Bengaluru, Karnataka, India · Hybrid

• Currently working in the areas of Object Detection and Tracking with the Computer Vision team.
• Highlights:
• ❏ Dec'23: Recognized with Team Award
• ❏ Sep'23: Recognized with Excellence Award
• ❏ Jul'23: Work selected (32/682 submissions) in SparkTech Summit 2023: People Unified Movement Analytics
• ❏ May'23: Won "Best technical achievement in a demo" Award for presenting work on Multi Camera Person Trajectory Estimation
• ❏ Apr'23: Work selected in Walmart AI Summit 2023: Estimating impressions for TV wall    ​
• advertisements via Computer Vision ​
• ❏ Oct'22: Recognized with Bravo Award
• ❏ Apr'22: Presented work in Walmart AI Summit 2022: Object-Detection and Tracking in Large-Scale Retail Environments
• ❏ Oct'21: Recognized with Associate of the Month Award
• Initiatives taken:
• ❏ Built DataTag: Auto-Labelling Tool for Text Classification
• ❏ Participated in Technical Debate on Explainability AI
• ❏ Built Chatbot (RAG) on Confluence pages using LLaMA 2-13B, FAISS and LangChain
• ❏ Built Cross-Modal Clothing Item Retriever using Contrastive Learning

Jan 2021 – Jun 2021 · 5 mos · Bengaluru, Karnataka, India

• Team: Consumer Payments Japan, Data Science
• Project: Customer2Vec Embeddings
• ❏ Borrowing concept from Word2Vec, Customer2Vec Embeddings are low-dimensional representations of customer data that capture customers' behaviour from a payments perspective
• ❏ Learned embeddings by designing a Neural Network to predict next month’s GMS (Gross Merchandise Sales) on data of 22.5 MM customers
• Project: Payment Migration and GMS Penetration Dashboard in AWS QuickSight
• ❏ Built two end-to-end business intelligence solutions from extracting data from Amazon Redshift to productionizing them at scale
• ❏ Presented dashboards to multiple teams from Finance, Marketing and Product, and received appreciation from senior leadership

Jun 2020 – Jul 2020 · 1 mo · Kolkata, West Bengal, India

• Project: Fantasy Games – Skill v/s Chance
• Advisor: Professor Sahadeb Sarkar
• In this project, we looked at different approaches that can be taken to identify whether a cricket fantasy game is skill-based or chance-based. I had simulated behavior of bots according to rules of game with Monte - Carlo simulations and compared payoffs, winning probability with that of actual players using statistical tests of means and distributions. I had also taken a Bayesian approach to validate my results.

Apr 2020 – Jul 2020 · 3 mos · Kharagpur, West Bengal, India

• Project 1: Explainable Recommendation Systems
• Advisor: Professor Niloy Ganguly
• For this project, we are interested in explaining why set of movies is recommended to users. The dataset that I had worked on is MovieLens 1M dataset. I started by constructing a graph where users and movies are nodes while the ratings given by users to movies are edges. I then calculated Personalized PageRank (PPR) scores for each user for different movies based on which recommendations are made for users.
• I had also embedded additional information about directors, actors and other movie attributes to enrich the knowledge graph by scraping information from DBPedia. After building an efficient recommendation model, I generated counter-factual explanations for recommendations for users using polynomial-time optimal algorithm.
• Project 2: Knowledge aware chatbots
• Advisor: Professor Pawan Goyal
• Normally, chatbots respond keeping the ongoing context in mind built by the interaction facilitated with user. However, for this project we are interested in building a chatbot where it can answer to queries information of which was fed a while back, and so it needs to have a memory from which it can continuously read and update. Also, the chatbot needs to keep learning from the information fed i.e. interactions in time, and so it’s continual/life-long learning.
• I had built a Memory Transformer modifying the Transformer architecture by adding a memory block. The dataset that I had used was Wizard of Wikipedia and the problem was modelled as reinforcement learning problem. The code is written in PyTorch.

Jun 2018 – May 2019 · 11 mos · Mumbai Area, India