May 2022 – Aug 2022 · 3 mos · Menlo Park, California, United States

• PyTorch Architecture Optimization Team
• Developed data sparsification API that allows users to prune trainable and non-trainable tensors.
• Added scheduling support to control sparsity levels during training. Built callbacks to support Pytorch lightning.
• Developed activation sparsifier API that allows users to prune activations flowing through the model.
• Integrated the data sparsifier API with quantization workflows.
• Used the data sparsifier API on the Deep Learning Recommendation Model to prune embedding tables. This compressed the model size by 80% with less than 2% loss in accuracy.

Jan 2022 – May 2022 · 4 mos

• Researcher at the CILVR Lab advised by Prof. Sumit Chopra
• Novel end-end deep learning framework that potentially reduces the screening time of an MRI exam by 20x
• Submitted paper to NeurIPS 2022 for review.

Jun 2018 – Jul 2021 · 3 yrs 1 mo · Bengaluru Area, India

• Majorly worked with a retailer chain wherein I built models to improve customer acquisition and engagement.
• Customer Cadence Model | developed on 2M customers
• Developed a model that classifies customers who are at risk of not returning to buy replenishable products.
• Different models were combined to balance precision & recall.
• In an A/B test setup, the model improved the conversion rate by 2% & the dollar sales by $2k/week.
• Product Recommender for Email Campaigns | built on 25M customers
• Developed the engine using implicit collaborative filtering and created a multiplier to prioritize seasonal items.
• Developed a framework to map customers to a set of offers based on the products recommended.
• In an A/B test setup, the model improved the dollar sales by 7% ($100k/month).
• Model Deployment
• Automated the cadence and recommender models to run in batch-mode using Ansible. AWS EC2 was used for computation and S3 for model & data storage. Used Snowflake to prepare the modeling data.
• Developed and deployed a lead-scoring model in real-time on Kubernetes. FastAPI was used to create the endpoint service.
• I also worked with a health division CPG Company:
• Developed a demand forecasting model.
• Scraped and created geographical features using OpenStreetMaps to improve model prediction

May 2017 – Jul 2017 · 2 mos · Tokyo, Japan

• Developed facility recommendation engine (for an internal product) using collaborative filtering using Spark.
• Used RedisDB in the backend to fetch and save recommendations

May 2016 – Jul 2016 · 2 mos · Bengaluru Area, India

• License Plate Recognition System.
• Built image processing pipeline to process and segment characters from raw license plate images.
• Used Convolutional Neural Network to recognize characters on License Plate.

Jun 2015 – Jul 2015 · 1 mo · Bangalore

• Embedded WEKA in an android app to perform a prediction task.