George Fraser

George Fraser is co-founder of Fivetran and chief executive of Fivetran + dbt Labs, the combined company formed in June 2026. He helped turn connectors built for a pivot-table product into a managed data-pipeline business. His more recent work connects that infrastructure with a question raised by AI analytics: how does an agent learn what a company’s data means, alongside learning how to query it?

From neural signals to data pipelines

George trained in neurobiology at the University of Pittsburgh. His research included a 2009 brain–computer interface paper, coauthored with Steven M. Chase, Andrew Whitford, and Andrew B. Schwartz, that tested whether decoding movement required first identifying which individual neurons produced recorded electrical spikes. The team extracted control signals without that sorting step; two rhesus monkeys achieved cursor control comparable to methods using sorted spikes. The result demonstrated a simpler processing approach for a neuroprosthetic interface.

Before Fivetran, George also worked as a research scientist at a biotechnology company, where he used MySQL to track metadata while keeping the underlying data in files. He later drew on that experience in his writing about databases and machine learning: scientists working with files and specialized libraries have different needs from teams querying relational data for business reporting.

Fivetran’s defining product emerged from a problem its founders initially treated as incidental. George and co-founder Taylor Brown were building a pivot-table interface for large datasets on Amazon Redshift. In 2014, the team, which included engineering leader Meel Velliste, built connectors to bring application data into the product’s backend. Customers wanted those connectors independently of the interface. When one offered to pay for a Salesforce-to-Redshift connection, the founders redirected the business toward data pipelines.

That pivot made ongoing maintenance the product’s central engineering problem. The team needed connectors that could accommodate changing APIs and database schemas and deliver data analysts could query. A managed service would absorb work that customers otherwise repeated across bespoke pipelines. George envisioned Fivetran becoming infrastructure beneath other products, supplying data even when their users did not know it was involved.

George subsequently pursued broader coverage of enterprise data. In his 2021 announcement of the agreement to acquire HVR, he described complementary expertise: Fivetran understood the intricacies of SaaS applications, while HVR specialized in replicating mission-critical databases with low latency and a light operational footprint. The proposed combination extended the company’s reach from application connectors to the systems holding large enterprises’ core records.

Architecture should follow the work

George’s technical writing examines architectural choices through the workloads they serve. His positions span how teams share data, how databases should be compared, and what inexpensive software changes about analytics:

• Interoperability across working practices. George distinguishes SQL-centered analytics teams from scientists whose work revolves around Python libraries and files. For teams sharing relational data across both workflows, he identifies Apache Arrow and shared lakehouse formats as ways to avoid duplicated effort. His argument leaves room for several architectures because the useful choice depends on how people work with the data.
• Benchmarks that test the relevant alternative. In his 2023 DuckDB experiments, George compared the embedded database with a commercial analytical warehouse. DuckDB performed strongly on smaller configurations, while the commercial system scaled better as data and CPU counts increased. His MacBook Pro also outperformed a tested 16-core server, leading him to advocate taking client-side analytics seriously. The results concerned those configurations, rather than establishing a permanent ranking of database engines; he also disclosed a small investment in MotherDuck.
• Disposable software built on maintained context. George uses “disposable software” for narrowly useful programs that can be created quickly and discarded. His reporting agent, assembled with Codex, combines descriptions of Fivetran’s dbt models with utilities for querying the warehouse. It produces reports through Python scripts, SQL, and Markdown. During a board meeting, he used it to revise two reports and generate another as the discussion developed. He expects analytics teams to put more effort into maintaining trustworthy models and context, alongside a smaller set of carefully maintained dashboards.

Bringing data movement and meaning together

The dbt Labs merger brings data movement and transformation into the same company, with George continuing as CEO and dbt co-founder Tristan Handy becoming president. George’s case for the combination emphasizes openness and interoperability: the platform should work across the surrounding ecosystem while AI applications continue to evolve.

His reporting agent illustrates why transformation and context matter to that ambition. Its warehouse-querying utilities give it access to data; descriptions of the dbt models help it interpret that data in the company’s terms. In George’s account, making reports cheap to generate increases the value of maintaining the definitions and models on which those reports depend.