Senior AI / Machine Learning Engineer

Worked on UpToDate Expert AI, a multi-agent clinical decision-support platform for answering complex medical questions using trusted clinical evidence, reasoning, and expert validation. Designed a LangGraph orchestration framework using Phi-3 Mini for intent classification, task decomposition, agent selection, tool routing, and workflow control, with escalation to larger models through Amazon Bedrock.

Built hybrid RAG and GraphRAG capabilities using PostgreSQL with pgvector, BM25, dense vector search, Reciprocal Rank Fusion, and Neo4j. Developed multimodal clinical evidence processing, fine-tuned a Llama 3.1 clinical reasoning model using PyTorch and LoRA, created MCP-based clinical tools, and implemented Bedrock Guardrails and DeBERTa-based safeguards. Deployed containerized services on Amazon ECS and built RAGAS-based evaluation and CloudWatch monitoring.

Contributed to TrialLens, an AI-powered clinical-trial analytics platform supporting connected medical devices, patient adherence, device synchronization, digital biomarkers, and eCOA data in decentralized clinical trials. Designed real-time streaming architecture on Azure with Event Hubs for device telemetry ingestion and Service Bus for reliable asynchronous processing and dead-letter handling.

Built production streaming pipelines with Databricks Auto Loader, Spark Structured Streaming, Delta Live Tables, Delta Lake, ADLS Gen2, Unity Catalog, and Databricks Workflows. Developed Python, Pydantic, PySpark, and SciPy validation and feature-engineering pipelines, low-latency analytics with Azure Data Explorer, and governed conversational agents using Semantic Kernel and Azure OpenAI Service.

Developed model-training and production inference frameworks for automated English scoring across pronunciation, fluency, vocabulary, grammar, prosody, timing, and discourse. Built Python pipelines with scikit-learn, SKLL, RSMTool, pandas, and NumPy for preprocessing, LASSO feature selection, hyperparameter tuning, validation, model comparison, and packaging using over 239,000 human-rated responses.

Designed an Apache Storm and Java runtime framework using reusable spouts, bolts, and topologies for ingestion, feature extraction, filtering, inference, and score aggregation. Deployed containerized Python and Java services on AWS using EC2, S3, ECR, and CloudWatch, scaling throughput to more than 12,000 scored responses per hour. Also developed quality-control models and a hybrid ML-plus-LLM scoring architecture with confidence-weighted score fusion.

Worked on customer behavior and churn analytics for T-Mobile, analyzing usage, billing, subscription, device, and campaign data to identify customers at risk of leaving and support targeted retention campaigns. Built a Kafka-based customer data pipeline with raw data stored in Amazon S3 and developed PySpark pipelines on Amazon EMR for large-scale telecom data processing and feature creation.

Created a central analytics dataset in Amazon Redshift and built an XGBoost churn prediction model in Python. Evaluated the model using ROC-AUC, precision, recall, and lift, and used Spark GraphFrames to create social-network features including PageRank, connection counts, communities, and neighboring churn behavior. Managed training and scoring workflows with Amazon SageMaker.