CHAPPIDI SASI

Designed and implemented end-to-end ML training and inference pipelines optimized for GPU-accelerated environments, supporting large-scale deep learning workloads across multiple teams. Built and optimized distributed training workflows using PyTorch, CUDA, and NCCL, improving multi-GPU training efficiency and reducing model training time by 30%+. Developed high-performance data preprocessing and feature pipelines using Python, Spark, and custom CUDA kernels, enabling TB-scale dataset processing. Collaborated with research scientists to productionize state-of-the-art transformer- and vision-based models, ensuring reproducibility and scalability. Integrated models into real-time and batch inference systems using Triton Inference Server and TensorRT, achieving low-latency inference. Optimized model performance through mixed-precision training, model parallelism, and quantization, reducing GPU memory usage by 20–25%. Built and maintained MLOps workflows for experiment tracking, model ve

Developed and deployed machine learning models supporting fraud detection, risk scoring, and user personalization across Paytm’s payments platform. Built end-to-end ML pipelines for data ingestion, feature engineering, training, and inference on billions of transaction records. Implemented supervised and unsupervised models, improving fraud detection accuracy by 28% while reducing false positives by 20%. Designed real-time inference services handling millions of transactions per day with low latency and high reliability. Optimized feature stores and data pipelines to reduce model inference latency by 30% under peak traffic conditions. Collaborated with product, risk, and data engineering teams to translate business requirements into deployable ML solutions. Implemented model monitoring, drift detection, and automated retraining, improving long-term model stability and performance. Deployed ML models using Docker and Kubernetes, enabling scalable and fault-tolerant production inferenc