Jay Shah

Led foundation-model MLOps platform; cut release time 45% and increased model adoption. Architected agent platform with MCP tools to spin up domain agents in under 5 minutes; now powering 20+ workflows. Built LLM interface integrating asset data, enabling reporting and task automation. Launched cross-site solar-storage autoencoder anomaly detection to predict failures and performance issues, saving >$500k. Shipped RAG compliance agent, reducing violations 10% and automating audit preparation for energy domains. Reduced false positives 25% by ranking predictive alerts and next-best actions with Bayesian analysis.

Deployed XGBoost-based solar/wind forecasting service; lowered MAPE 10% over baseline. Mentored offshore ML team on systems design and best practices; boosted delivery speed 25%. Defined ML architecture standards adopted by 3 squads to streamline handoffs and deployments.

Built wind-turbine remaining useful life (RUL) model (92% accuracy) using industrial sensor and fault data. Automated model deployment and tracking using Airflow, Docker, Serverless, and Terraform on AWS/GCP. Designed Airflow + AWS Lambda ETL reducing latency 60% across five clients. Prototyped statistical tools to quantify power production inefficiency and energy loss.

Implemented anomaly detection to predict component failures using GBM for 8 wind-turbine components. Estimated bearing types and segmented bearing failures from 10-minute signatures using K-means clustering; delivered actionable insights.

Researched with Dr. Yu Ding on applying advanced machine learning methods to solve and predict wind energy system failure. Implemented deep learning methods to predict possible power production and downtimes associated with the failures of wind turbine.

Created weather-controlled building baseline regression models for all digitally metered utilities using enterprise energy module. These models are used to monitor consumption across the campus to prevent sensor issues and energy loss. Manipulated Data in SQL to compare baseline modelled consumption with real-time consumption using statistical control limit chart in excel to analyse the average variation related to prediction.

As a Data Ambassador for DataKind, I played a pivotal role in a collaborative project with John Jay College, developing advanced machine learning models to predict student dropout and delayed graduation. Leveraging Random Forest classifiers, our team crafted and tested over 20 models, ultimately recommending a tailored suite of six models to enable early identification and support for students at risk, significantly contributing to improving college completion rates.