Vaishnavi Kaleru

Designed and deployed a transformer-based Payments Foundation Model using self-supervised learning, fine-tuning, and RAG pipelines on billions of transactions, generating embeddings for fraud detection and authorization optimization. Built distributed training pipelines across GPU/TPU clusters on GCP Vertex AI and AWS SageMaker for scalable experimentation. Developed real-time inference microservices with FastAPI, Docker, Kubernetes (GKE & EKS), ONNX Runtime, TorchScript, and Triton Inference Server, reducing latency for fraud scoring. Implemented end-to-end feature pipelines with Spark, PySpark, Airflow, Kafka, Redis Streams, and feature stores for low-latency ML inference. Applied prompt engineering, QLoRA fine-tuning, and multimodal fusion to domain-specific payment data, improving fraud detection accuracy by ~64%. Implemented MLOps pipelines with MLflow, Weights & Biases, Terraform, ArgoCD, and Azure DevOps, including CI/CD, model versioning, canary deployments, rollback strategies

Designed end-to-end AI pipelines for real-time analysis of music and performance videos using Python, PyTorch, Hugging Face Transformers, and transfer learning with CNN/Transformer backbones, achieving >98% accuracy in detecting musical patterns, rhythm, and performer gestures. Built and maintained large-scale audio/video data pipelines with Python, Apache Spark, Airflow, and Hive/Presto, automating preprocessing for 50,000+ hours of performance recordings. Implemented distributed training workflows on Azure GPU clusters with Docker, Kubernetes, and TorchElastic, reducing model training time by 35% while maintaining state-of-the-art performance across diverse genres. Optimized production-ready multimodal models using ONNX, TensorRT, and Azure ML, achieving sub-100ms latency for real-time music analysis and recommendations on edge/mobile. Applied ML lifecycle management with MLflow, Weights & Biases, and Optuna, enabling reproducible results, hyperparameter tuning, and efficient iterati