Public: Ramp Research agentic data analyst

Architectural notes

Per Ramp's engineering blog "Meet Ramp Research" and the ZenML LLMOps
writeup, Ramp Research is an agentic data analyst that answers questions in Slack;
Ramp reports ~1,800 questions across ~1,200 conversations in four weeks.

Ramp aggregates and indexes metadata from Snowflake (warehouse), dbt, and Looker,
plus domain docs organized as a filesystem the agent navigates via tool calls;
these data systems do not map to catalog component slugs and are noted here.

It uses agentic tool-calling to inspect column values, branch exploration paths,
and backtrack, rather than single-pass RAG; no vector database or embeddings are
described.

Evaluation is a deterministic Python mini-framework built into their dbt project
that validates expected tool calls, table references, and query structure
(in-house, represented as eval: local).

Access controls exclude PII at the table/column level. Not publicly disclosed: the
underlying LLM model/provider or the agent framework.