AI/BI Dashboards
A low-code dashboarding product for Lakehouse data, launched in 2023.
Justin Talbot
Principal Software Engineer · Databricks
I’m a Principal Software Engineer at Databricks, where I work on AI/BI — systems that let users answer their data questions through natural language and visual interfaces.
I earned my PhD in Computer Science from Stanford University. My research spans data analytics systems and information visualization: compilers and runtimes for vector and array languages, automatic schema inference, algorithms for the automated graphical design of visualizations, and the graphical perception experiments that tell us which designs actually work. Earlier, my master’s work in computer graphics introduced resampled importance sampling, which has since become the foundation for ReSTIR and related advances in real-time rendering.
Work
Databricks 2021 – present
Unity Catalog business semantics
A governed semantic layer for business metrics. Metric views attach measure definitions to SQL relations, bringing metrics into a single relational model rather than a separate layer alongside it.
Apache Ossie
An open, vendor-neutral standard for metric and dimension definitions, incubating at the Apache Software Foundation.
Tableau 2012 – 2020
Level of detail expressions
LOD expressions let a calculation aggregate at a level of detail independent of the view, which had previously required restructuring the data. Grew out of an internal research project and shipped in Tableau 9.0 in 2015.
Data model and relationships
A logical modeling layer above physical joins, shipped in Tableau 2020.2. Relationships resolve the right aggregation level per view rather than forcing every question through a single join grain.
Query optimization
Query fusion for complex analytical queries, reducing the number of round trips a single view generates.
Schema inference
Deriving a normalized schema from a denormalized table, framed as an unsupervised learning problem.
Stanford University PhD, 2013
Riposte
A trace-driven compiler and parallel virtual machine for R. Riposte records vector operations at runtime, fuses them into compiled traces, and executes those traces in parallel, reaching performance competitive with hand-written C on unmodified R code.