From prompt to PR: how Careem’s AI writes and validates dbt models before a human ever sees them

Careem, a UAE-based super app with more than 5,000 dbt models supported by a lean in-house data engineering team, developed an AI-driven framework to reduce a backlog of data requests that previously required one to two weeks each. Rather than relying solely on LLM-generated code, the platform emphasizes end-to-end validation through a four-part architecture that supplies metadata and business context, requires clarification of ambiguous requirements, separates governed metrics from temporary analyst outputs, and isolates AI changes in sandboxed environments. The agent automatically compiles, tests, materializes, and corrects dbt models before opening a pull request for human review, including checks for data types, source-table selection, data movement across layers, and domain-specific storage conventions. In a demonstrated request for daily merchant food revenue, the system resolved two test failures independently before producing a documented, tested model. Careem reports that the approach has reduced metric delivery from weeks to hours, now generates roughly 90% of new models, has caused no highest-severity incidents, prevents duplicate requests through dbt manifest checks, and helped remove 3.6 petabytes of temporary data through automated lifecycle management.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.