Architected a multi-agent conversational AI framework for real-time English tutoring across text and voice channels, applying ReAct and Chain-of-Thought reasoning with MCP to ensure consistent state handling and memory alignment across agents. Implemented hybrid Knowledge Retrieval using Pinecone vector search and GraphRAG (Neo4j) to surface cultural nuances and context-aware prompts, improving response accuracy and learner comprehension. Built and optimized a PEFT fine-tuning pipeline (QLoRA, Adapters, Instruction Tuning) on GPT-4 / Llama models through LangGraph, reducing training cycles and hallucinations. Deployed scalable MLOps infrastructure with Terraform + AWS EKS/ECR, increasing peak session throughput 3×. Developed an Automated Tutoring Quality Evaluation Suite to measure relevance, tone, and cultural sensitivity, enabling continuous A/B iteration on tutor persona. Implemented RLHF feedback loops (PPO/GRPO) to prevent model drift. Optimized inference latency using vLLM, Tens

Data Scientist at Sage

August 1, 2015 - May 1, 2022

Built advanced NLP pipelines on Databricks and Snowflake for financial data, including a BERT-based entity recognition system with spaCy in TensorFlow, reducing data extraction error rates from 11% to 2%. Developed financial multi-agent chat systems using LLM-based orchestration (LangChain Agents) and adaptive context-aware AI agents, slashing human intervention by 60% and improving workflow efficiency. Created an LLM-powered summarization platform for real-time financial news using GPT, NLTK, and PySpark, reducing analysts’ research time by 75%. Implemented a sentiment analysis engine for investment insights (PyTorch, GCP AutoML), achieving 92% accuracy on 500k+ articles. Orchestrated end-to-end MLOps with Kubeflow, Terraform, and CloudFormation, expediting deployments by 30% while maintaining automated LLM testing for stable releases.

Machine Learning Engineer at IBM

March 1, 2013 - July 1, 2015

Engineered an automated OCR pipeline (Tesseract, OpenCV, Keras) for client onboarding, reducing manual data entry effort from 100,000 to 30,000 hours annually. Boosted product engagement through a recommendation engine (scikit-learn, PyTorch) for e-commerce personalization. Deployed a healthcare document classification system (TensorFlow, NLTK), achieving 97% accuracy on ~1,000 daily records. Developed time-series forecasting models (ARIMA, LSTM, GRU) to detect risk patterns and forecast trends. Created executive Tableau dashboards for real-time customer behavior insights, guiding data-driven strategies and improving campaign ROI.