Nikki Potla

Developed production-grade AI agents and enterprise knowledge assistants using Python, LangGraph, MLflow 3.0, Databricks Vector Search, and Unity Catalog. Improved response accuracy through automated LLM evaluation pipelines, synthetic data generation, prompt and retrieval optimization, document chunking, and continuous benchmarking. Reduced inference cost and latency via intelligent model routing, retrieval optimization, caching, and Databricks Model Serving. Built secure, governed RAG platforms across 50M+ documents using Delta Lake and Unity Catalog, with evaluation, tracing, and observability using MLflow Evaluation/Tracing and Databricks SQL. Architected cloud-native AI microservices with FastAPI and Kubernetes, deployed on AWS with high availability, and automated CI/CD and infrastructure using GitHub Actions, Terraform, and Databricks Asset Bundles. Engineered distributed pipelines with Spark Structured Streaming and Delta Live Tables to process 2B+ records annually.

Built low-latency LLM inference services using TensorRT-LLM, Triton Inference Server, PyTorch, and CUDA, improving throughput and reducing costs through quantization, dynamic batching, KV cache optimization, and request scheduling. Developed inference APIs with REST/gRPC and frameworks such as vLLM to support RAG and agentic workloads while maintaining sub-second latency and improved accuracy. Engineered high-performance model serving for GPU clusters using TensorRT-LLM, Triton, CUDA, NCCL, and Kubernetes, optimizing distributed inference with NCCL-based communication, memory optimizations, and intelligent routing. Implemented OpenAI-compatible REST/gRPC services and deployed on AWS EKS with Prometheus/Grafana monitoring. Automated CI/CD with GitHub Actions and infrastructure automation for reliable enterprise deployments and governance.

Developed Python and PySpark demand forecasting pipelines on Databricks processing 50M+ supply chain records daily. Improved forecast accuracy by 22% using XGBoost, LightGBM, Prophet, and feature engineering; built scalable ETL and feature engineering pipelines with Spark SQL, AWS Glue, and Delta Lake integrating ERP and external datasets. Designed FastAPI microservices for forecasting, anomaly detection, and inventory optimization, reducing prediction latency by 35%. Implemented end-to-end MLOps with MLflow, Docker, Kubernetes, AWS SageMaker, and CI/CD for training, deployment, monitoring, and model versioning. Evaluated and selected approaches based on business requirements, scalability, and long-term maintainability, and collaborated with stakeholders to deliver dashboards and human-in-the-loop workflows supporting 4,000+ users and reducing stock-out incidents by 18%.