Sakthivel N.

Architected a production LLM assistant utilizing Graph-based RAG (LangGraph + Amazon Bedrock), enhancing query response relevance by 35% compared to baseline vector search. Designed an Orchestrator and Memory agent for multi-step reasoning over financial documents, significantly reducing analyst report turnaround from 4 hours to 30 minutes. Implemented sub-second real-time inference with Spring Boot APIs, handling over 500 financial queries per day, leading to a 45% decrease in manual data-analysis efforts. Developed and deployed a lead scoring model using Scikit-learn and TabNet on over 200K records to enhance sales prioritization efficiency by 25% while reaching an ROC-AUC of 0.82. Utilized SMOTE, K-Means clustering, and pseudo-labeling to improve model recall by 15% on imbalanced data, and lowered the model retraining cycle time by 25% through a modular, automated feature-engineering pipeline. Managed the cloud migration of legacy IVR workflows to Amazon Connect, achieving a 20% reduction in system downtime and an enhancement in production uptime SLA to 99.5% across three client environments. Conducted thorough end-to-end testing and validation, ensuring a zero-defect production cutover while eliminating on-premise bottlenecks.