Kavay Khurana

I’m a Data Scientist (I) at Meesho, focused on search personalization, ranking, and large-scale machine learning pipelines. I’ve worked on PRISMv2, a production-grade search personalisation and ranking pipeline on Databricks.

I contributed to training a multitask deep learning model (FT-Transformer) to predict pCTR, pCVR (order given click), and pCTCVR, delivering NDCG@1 improvement of +9% and NDCG@50 improvement of +6% over the positional baseline. I also architected an end-to-end PySpark feature engineering pipeline (15+ feature groups) with serialized Spark ML Pipeline to ensure zero train-inference skew.

Prior to that, I interned as an SDE at Bain and Company, where I built and deployed an internal AI assistant using Custom GPT integrated with Microsoft Graph APIs. Using Azure Functions, I engineered the backend integration layer for authentication, request routing, and scalable access to Microsoft 365 workflows via natural language.

My project work reflects my deep learning and experimentation mindset—such as a two-tower neural recommendation system in PyTorch (Transformer-based user tower, tracked with MLflow) and a speech command recognition model with TDNN and Conv1D layers (improved via dataset augmentation and fine-tuning). I’ve also placed 17th among 20,000 teams in Amazon’s India-wide ML Challenge and won ACM HACKLIPSE 3.0 among 100+ teams.