At Swiggy, I build a unified user-embedding foundation model for personalization using two-tower retrieval and BERT-style transformer encoders. I own the lifecycle from Spark and Databricks data preparation through distributed PyTorch training, ElastiCache feature-store serving, monitoring, and deployment validation.
I've also built RAG and OCR pipelines with AWS Bedrock, benchmarked ONNX models on NVIDIA Triton and vLLM, and optimized infrastructure, Gurobi discount models, and Snowflake writes. Earlier, I deployed multimodal, forecasting, semantic-search, sentiment-analysis, and computer-vision projects across BlueBagels, Innomatics Research Labs, and RGUKT.
