Guided tour
Marketing Warehouse. Ten sources in one place a model can read.
A BigQuery warehouse that pulls ten marketing sources into one place every morning and joins ad spend through to booked patients, with a catalog an LLM reads and a weekly graded eval.
Walk through it, six steps
Ten sources, one morning sync
Cloud Scheduler fires the connector jobs at 6 AM: ads, calls, CRM, web, search and social, with one platform arriving via BigQuery Data Transfer. All connectors share retry, load, and secret code, packaged as a single container image run as serverless jobs.
| Source | Rows | Status | Finished |
|---|---|---|---|
| Search ads (Platform A) | 1,204 | ok | 06:04 |
| Social ads (Platform B) | 912 | ok | 06:06 |
| Call tracking | 388 | ok | 06:07 |
| CRM boards | 2,730 | ok | 06:11 |
| Search ads (transfer) | · | managed transfer | 06:15 |
The problem.
Marketing questions like which channel actually produces patients required manual exports from each platform and stitched spreadsheets. And once an LLM was allowed to answer questions over the data, a harder problem appeared: how do you know its answers stay right as the schema evolves?
- 10 ads, calls, CRM, web, search and social
- 98.5% patients joined to leads; 99.3% agreement with the weekly exports
- 22 golden Qs weekly graded run, fails under 80%
- 6:45 AM freshness and drift check every morning
How it fits together.
Ingest
- 10 sync jobs, one Docker image, Cloud Run jobs
- Shared retry, load and secret modules
- 6 AM schedule + managed transfer
Warehouse
- BigQuery: raw → staging → marts (+ evals dataset)
- Inline descriptions on every view and column
- data_guide catalog written for LLM readers
Consume
- LLM access via MCP Toolbox
- Looker Studio dashboards
- Plain-English answers with mart citations
Trust
- 6:45 freshness check + z-score drift sentinel
- 22 golden questions graded weekly, 80% gate
- Glass monitoring app with per-node health
Where it landed.
Marketing spend, calls, leads and booked patients live in one queryable place: 98.5% of patients link to a lead, with 99.3% agreement against the weekly exports. The part worth stealing: treating the LLM's access as a product surface with its own weekly, graded, gated eval suite instead of assuming yesterday's good answers persist. All dollar figures shown are synthetic.