Built supervisor-based agent workflows that generate structured JSON for dynamic security dashboards. Designed outputs with data endpoints, UI components, and layouts, enforced strict prompt rules for stability, and used RAG to pull schema and component metadata. Deployed on AWS Lambda with CloudWatch monitoring and partnered with platform teams to harden Kubernetes pipelines.

Built agentic AI workflows with LangChain/LangGraph using Claude Sonnet 4 to automate support tasks. Added persistent state to multi-agent flows to cut repeated interactions by 30%+, built supervisor–worker patterns to reduce escalations, integrated MCP tools for internal lookups, and shipped to AWS Lambda with CloudWatch and Kubernetes-driven CI/CD.

Built an XGBoost forecasting model trained on 4.2 million advertising records, improving weekly ROI prediction accuracy by 15 percent and supporting multi-million-dollar budget decisions.

Developed PySpark ETL pipelines processing over 300,000 daily records, cutting feature engineering time by 40 percent and enabling daily retraining.

Deployed real-time inference endpoints on SageMaker with automated scaling, achieving sub-120ms P95 latency and reducing infrastructure costs by 30 percent.

Optimized Flask APIs by restructuring SQLAlchemy queries and adding indexing, improving throughput by 30 percent and reducing report load times from 12 seconds to under eight.

Implemented experiment tracking with DVC to strengthen reproducibility and collaboration across data teams.