Shrink the surface: governed self-service for analyst-owned dbt pipelines

At a dbt Summit 2026 breakout session titled "Shrink the surface: governed self-service for analyst-owned dbt pipelines," Jinze Xin, analytics engineer at Liftoff, described how narrowing what analysts are responsible for, rather than teaching them more of the platform, finally made self-service dbt stick on his team.

Liftoff is an AI-powered marketing and monetization platform for the mobile app economy, helping advertisers, publishers, and game developers reach and monetize mobile users. Xin cited public filing figures showing the platform's SDK reached more than 1.4 billion daily active users in the first quarter of 2026, across more than 167,000 integrated apps.

Before the project she described, Liftoff's roughly ten-person supply analytics team faced three recurring problems. Every new production dataset meant a ticket into an engineering queue where someone had to reconstruct the analyst's business intent from scratch, a round trip that could take weeks; recurring questions were answered by rescanning billions of rows of raw event data instead of a curated dataset; and ad hoc requests pulled engineers off their own roadmap work.

Liftoff's first attempt at fixing this was straightforward: give analysts direct access to the production dbt project, train them on dbt concepts, write documentation, and offer engineering support. Adoption never stuck, because access alone didn't change how much an analyst had to learn, including the repository structure, approved sources, git workflow, and concepts like incremental strategy.

Xin's reframing was that analysts and dbt pipelines describe the same dataset in two different languages, and the first self-service attempt had simply asked analysts to do the translation themselves. The fix was to build a translation layer that could preserve an analyst's business intent while still producing an explicit, reviewable production contract.

That translation layer became dbt-pico, a separate dbt project wired to Liftoff's main engineering-owned project through dbt Mesh cross-project references. Analysts working in dbt-pico own only business logic and metric grain; the platform owns incremental strategy, CI, scheduling, review paths, and deployment.

An AI agent built on Claude Code skills is the entry point to dbt-pico. Analysts describe what they need in business language, and the agent grounds that request in governed metadata and lineage, checks whether an existing dataset already answers the question, and only then translates the remaining request into a dataset contract that carries both business meaning and engineering requirements.

The first version of that agent worked, but it produced too much: one early request touched 11 files, added roughly 620 lines, and introduced five intermediate models, a custom macro, and a project-level configuration change. Xin's team responded by shrinking the target output itself, consolidating the pattern down to a single shared macro that now carries incremental strategy, event-time handling, and ownership metadata for every model in the project, a change that removed 564 lines from the codebase.

Liftoff also constrained the agent along two separate dimensions. A "building surface" limits it to governed upstream references, a fixed model shape, and shared configuration, while a separate "authority surface" lets it discover context, propose a model, and open a pull request, but never merge code, run production compute, change a schedule, or trigger a backfill.

To manage hallucination risk, Liftoff paired the agent's skill-based guidance with deterministic checks: a CI governance script validates model naming, ownership, and required metadata against every changed file at no additional model cost. Two further validation paths, both requiring a human to trigger them, materialize sample data and compare output against expectations before a pull request merges.

The net result, in Xin's telling, is that the analyst-facing workflow now has three steps: describe the intent, review the evidence, and approve the meaning. An internal analytics lead described dbt-pico as "a step-function unlock for analysts," calling it the first time the team had been able to build first-class, self-serve data pipelines for supply-side work.

Xin shared early results: one production dataset that previously took around eight weeks to ship through the engineering queue went into production in under a week once an analyst built it through dbt-pico. A recurring analytical query that took roughly 30 minutes against raw event data now returns in under a minute against pre-aggregated datasets, and engineering time freed up by fewer ad hoc requests went back into planned project work between March and May.

Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →