Developed RAG based legal chatbot built on top of the chat service for a legal marketplace for finding lawyers.

Implemented counterfactual evaluation methods that can emulate real A/B test results from historical data, thus reducing reliance on A/B tests and saving thousands of hours monthly in iterative development time.

Adapted ideas from research papers for experiments on internal user data using GCP and BigQuery. Integrated the best methods in Snap’s in-house ML platform using Apache Beam, which outperformed existing metrics.

Developed Computer Vision models with Pytorch & Tensorflow for semantic and instance segmentation using weakly supervised techniques. Built efficient data pipelines for model inference on gigascale images on cloud.

Deployed product on AWS using EC2, SageMaker, Lambda, S3, SQS, DynamoDB etc.