Architected and optimized large-scale deep learning training and inference pipelines using PyTorch and Meta's FBLearner, processing billions of data points daily to power content ranking and recommendations across Facebook and Instagram, contributing to double-digit improvements in engagement metrics. Built and deployed production Retrieval-Augmented Generation (RAG) and Graph-RAG pipelines leveraging LLaMA models and vector databases (Pinecone, ChromaDB, FAISS), improving contextual response accuracy by 30% for AI-powered product features serving 100M+ active users. Fine-tuned foundation LLMs using LoRA, QLoRA, and PEFT techniques combined with RLHF on domain-specific datasets, reducing inference compute costs by 25% while maintaining task performance benchmarks across production systems. Designed multi-step agentic AI workflows using LangGraph, AutoGen, and Model Context Protocol (MCP), enabling autonomous reasoning, tool use, and multi-turn planning capabilities embedded in Meta AI

Machine Learning Engineer at Accenture

March 1, 2022 - June 1, 2024

Developed and deployed 10+ production ML models using Python, TensorFlow, and Scikit-learn across enterprise client environments, with end-to-end CI/CD pipelines on AWS and Azure, consistently exceeding accuracy benchmarks and cutting model deployment timelines by 35%. Built scalable data preprocessing and feature engineering pipelines with Apache Spark and Pandas, processing millions of records while ensuring data quality standards for downstream model training. Engineered NLP classification and predictive analytics solutions including customer churn prediction, document classification, and fraud detection, optimizing hyperparameters via Optuna and GridSearchCV to maximize model performance. Integrated production-ready ML models into microservices architecture via FastAPI REST APIs, achieving sub-200ms inference latency and enabling seamless real-time integration with client applications. Established model monitoring and observability infrastructure using MLflow and Databricks, tracki