Shankar Nayak

About Shankar Nayak

I am a Senior AI/ML Engineer at Perplexity AI with 4+ years of experience delivering production-grade GenAI and MLOps solutions. My expertise lies in building Agentic AI and RAG architectures using LangGraph, CrewAI, and OpenAI to solve complex problems in healthcare and finance. I specialize in the full model lifecycle: from fine-tuning LLMs with PEFT (LoRA/QLoRA) to deploying scalable APIs via FastAPI, Docker, and Kubernetes. I have a proven track record of managing high-scale data with Spark and Kafka, and automating cloud infrastructure using Terraform and AWS CDK. With a B.Tech from IIT Guwahati and an MS in Data Science from Pace University, I combine deep technical research with an MLOps mindset to drive measurable business impact.

Experience

1. Machine Learning Engineer

   Perplexity

   Jan 2024 — Present

   Developed end-to-end machine learning pipelines using Python, TensorFlow, and PyTorch for processing 100M+ vehicle sensor data points daily, enabling predictive maintenance and autonomous driving feature development. Built real-time ML inference systems with FastAPI and AWS SageMaker, serving 50K+ predictions per minute with sub-100ms latency for battery optimization and vehicle performance monitoring. Implemented a computer vision perception module for autonomous vehicles, achieving 96% accuracy in object detection using CNN architectures and optimizing inference with OpenCV\'s UMat and ONNX Runtime. Designed advanced feature engineering workflows using Apache Spark and Kafka for real-time sensor data processing, reducing model training time by 40%. Applied LSTM networks for time-series forecasting of battery health, contributing to a 22% improvement in range estimation accuracy. Designed and implemented the end-to-end MLOps pipeline using MLflow and Kubeflow to automate the model lifecycle from training to deployment. Containerized models were deployed on AWS EKS via Docker and Kubernetes, reducing time-to-production by 45% while ensuring 99.9% uptime with auto-scaling. Pioneered a proof-of-concept RAG-based diagnostic assistant using LangChain and Hugging Face models, demonstrating a potential 50% reduction in technical information retrieval time. Collaborated with data science teams to translate automotive engineering requirements into scalable ML solutions, delivering 12+ production models.