About

I have worked in the capital markets industry for over 35 years — most recently designing end-to-end machine learning platforms and quantitative models for institutional investors. I hold an MS in Investment Management from Pace University, an MS in Analytics from the Georgia Institute of Technology, and am currently a PhD candidate in Computer Science, concentrating in Data Science, at Florida Atlantic University, where my research focuses on large-scale data systems, NLP, and reinforcement learning.

During my published research on predictive embedding frameworks (Albright & Zhu, Springer 2026) — processing over 3.3 million SEC Form 4 filings across 20 years to detect material semantic shifts in corporate insider trading behavior — I conducted the analysis by calculating vector distances with NumPy and scikit-learn across Parquet files, DuckDB queries, PostgreSQL tables, and .npy vector arrays, using JSON to track experiment results across runs. I initially avoided Apache Iceberg during the research: deploying Apache Spark was complete operational overkill for the workload, while PyIceberg did not feel mature enough for complex, high-throughput analytical pipelines.

Following the paper submission, the operational friction and severe data duplication of shuttling identical datasets across disconnected relational tables, raw array dumps, and search files led me to build BenoStreamDB. I designed BenoStreamDB to make Apache Iceberg dead simple and zero-friction to set up — running embedded with just a file URI or S3 path, without needing Spark clusters, external catalog services, or heavy JVM runtimes.

Starting initially as an in-process vector store directly over Iceberg tables to eliminate data duplication, I soon realized that unifying vector distance search, BM25 text retrieval, and tabular data in a single lakehouse table provided the exact foundation needed to derive and traverse relationship graphs in-place for Graph RAG — without operating an external graph database.