Statistical Rigor
Principled hypothesis testing, probability calibration, and continuous drift detection to ensure models remain dependable in production.
Staff Data Science Engineer Distributed ML & Streaming
Building sub-millisecond streaming feature stores, distributed lakehouse pipelines, and calibrated predictive models running in high-stakes production.
Data Engineering & Applied Science
Bridging the gap between mathematical inference and high-throughput production infrastructure. A machine learning model is only as dependable as the pipelines that feed it and the latency bounds it upholds.
My focus is end-to-end: formulate statistical loss surfaces, eliminate training-serving skew with declarative feature stores, and deploy quantized models into low-latency serving clusters with automated drift telemetry.
Principled hypothesis testing, probability calibration, and continuous drift detection to ensure models remain dependable in production.
High-throughput pipelines with Spark, Kafka, and Redis delivering sub-millisecond features to real-time inference endpoints.
Hermetic Docker environments, strict schema contracts with dbt, and automated continuous evaluation with Ray and Triton.
Selected work
Machine Learning · MLOps · 2025
A sub-15ms fraud detection pipeline processing 25,000 transactions/sec with a calibrated ensemble model, cutting chargebacks by 44%.
View case studyDistributed Systems · Data Engineering · 2025
A distributed streaming feature store ingesting 1.8B events/day with Apache Spark & Kafka, achieving p99 retrieval latency under 4ms.
View case study
Distributed Systems · MLOps · 2025
Sub-30ms hybrid lexical-vector retrieval architecture over 40M enterprise documents with semantic caching and cross-encoder reranking.
View case studyAll case studies illustrate production streaming architectures, benchmarks, and model systems.
Technical Capabilities
High-throughput, calibrated machine learning models designed for real-time inference.
Explore serviceReal-time event streaming and scalable Lakehouse data platforms.
Explore serviceRobust model registries, serving clusters, and automated continuous evaluation.
Explore serviceWorking principle
“In high-stakes production, an accurate model that misses SLA latency bounds is indistinguishable from a total system failure.”
From the journal
Engineering Note
How to design unified streaming feature ingestion with Kafka and Redis that satisfies sub-5ms p99 SLAs while eliminating training-serving skew.
Read articleWhy modern deep neural networks produce overconfident probabilities, and how temperature scaling and isotonic regression prevent catastrophic errors.
Read article
Architecture Guide
A hands-on guide to containerizing complex ML ensembles with NVIDIA Triton, dynamic batching, and Ray Serve concurrency controls.
Read article