Blobby
A data analysis agent built by Omni on top of Claude. It supports a complete pipeline that converts natural language into semantic queries and then into SQL, and incorporates an agentic loop with multi-step data analysis capabilities.
Timeline (last 90 days)
Omni used Claude to build a data analytics Agent called Blobby from scratch, going through 18 months of development.
Blobby's workflow translates user questions into Semantic Queries, then converts them into SQL, runs them in a data warehouse, and returns the results.
Blobby's first phase introduced three categories of metadata: AI context, sample queries, and values (field value examples).
The agentic loop is very good at recovering from errors, so Omni made Blobby aware of how to recover from errors, gave it a budget to try, and refined high-quality error messages.
As conversation complexity increased, Blobby's model was switched from Haiku to Sonnet, and token consumption rose accordingly.
Blobby experienced a "split brain" problem: the outer agent did not know which questions could be answered with a single query, and passing compound questions requiring multiple queries to the subagent caused confusion; the solution was to elevate the tools directly to the outer agent harness.
The current Blobby has a complete agentic system: the outer loop handles checkpointing and failure recovery, while the inner loop has a toolset for generating dashboards, generating visualizations, data modeling, and validation.
All Facts (7)
Omni used Claude to build a data analytics Agent called Blobby from scratch, going through 18 months of development.
50%UnverifiedBlobby's workflow translates user questions into Semantic Queries, then converts them into SQL, runs them in a data warehouse, and returns the results.
50%UnverifiedBlobby's first phase introduced three categories of metadata: AI context, sample queries, and values (field value examples).
50%UnverifiedThe agentic loop is very good at recovering from errors, so Omni made Blobby aware of how to recover from errors, gave it a budget to try, and refined high-quality error messages.
50%UnverifiedAs conversation complexity increased, Blobby's model was switched from Haiku to Sonnet, and token consumption rose accordingly.
50%UnverifiedBlobby experienced a "split brain" problem: the outer agent did not know which questions could be answered with a single query, and passing compound questions requiring multiple queries to the subagent caused confusion; the solution was to elevate the tools directly to the outer agent harness.
50%UnverifiedThe current Blobby has a complete agentic system: the outer loop handles checkpointing and failure recovery, while the inner loop has a toolset for generating dashboards, generating visualizations, data modeling, and validation.
50%