At Blue Yonder, I built TFX and Kubeflow pipelines for item- and location-level stockout risk forecasting, reaching 0.82 precision. I also developed alert-threshold tuning components so merchandising teams can balance precision and recall.
I developed probabilistic demand forecasts and a FastAPI scoring service for distribution-center stockouts, and rearchitected the transit lead-time model as a monotonic quantile network.
Previously, at Tiger Analytics, I implemented a RAG and Graph Neural Network system for SQL column extraction and built an invoice-intelligence pipeline. At RingCentral, I designed semantic search trackers and worked on meeting summarization; in research, I implemented Transformer-XL and proposed training methods that improved WikiText-103 perplexity.
