Pick a data role, then follow the graph

Turn messy tables into answers someone can act on. SQL, spreadsheets, a BI tool, and the nerve to say the number is wrong.

You ship: Trusted dashboards, ad-hoc SQL, and a written recommendation — not a model.

Own the serving layer: semantic models, certified datasets, and reports that finance will sign. Closer to the warehouse than a classic analyst.

You ship: Certified datasets, governed metrics, and reports that survive a refresh.

Move data that other people can trust: ingest, transform, orchestrate, and keep the warehouse bill from becoming a surprise.

You ship: Pipelines with tests, an on-call story, and a warehouse someone else can query.

The dbt-shaped role: models, tests, CI, and a metric finance can reuse. Between analyst questions and engineer pipelines.

You ship: Versioned marts, contracts, and a metric layer someone else can query.

Decide with data: experiments, models, and uncertainty. You still need SQL. You do not need a new architecture every sprint.

You ship: An experiment, a model with a holdout, and a decision memo — not a demo notebook.

Train, evaluate, and serve models as systems: features, latency, and rollback. Closer to production than research.

You ship: A served model with a metric, a rollback, and a feature table that matches training.

Make models a boring production service: CI, registries, feature stores, and the pager. Most of the pain is data platform pain.

You ship: A path from commit to served model with rollback, plus the monitors that catch drift.

Ship products on pretrained models: retrieval, eval, and guardrails. The hard parts are data, not the model brand.

You ship: A retrieval pipeline with evals, access control, and a rollback — not a chat demo.

Choose contracts: grain, ownership, lake vs warehouse, and what you will not build. You still need to have shipped.

You ship: A platform sketch the org can fund, plus the standards that keep it from rotting.