Case study 01

Replacing a CDP with a warehouse, a segmentation layer and an MCP server.

Café Zupas, 90+ locations / 2024 to present

$240K per year in license fees, gone.

The problem

The brand was paying a third-party customer data platform a quarter of a million dollars a year to do three things: stitch identity, build segments and hand audiences to Braze. The CDP owned the logic, the data was a black box to the analysts, and every new segment was a ticket to the vendor.

Architecture: sources to BigQuery warehouse to segmentation layer to Braze and to an MCP server used by ClaudePOS, loyalty, app, web and Braze events flow into hourly BigQuery marts. A SQL-defined segmentation layer compiles audiences and syncs them to Braze for email, SMS, push and in-app. An MCP server exposes the same marts to Claude under least-privilege IAM so employees can ask questions and see the SQL.POSLoyalty (SessionM)Mobile appWeb (GA4 export)Braze events01BigQuery warehouseHourly marts, mapped IDsMetric definitions in GitLeast-privilege IAM02Segmentation layerSQL segment catalogReal-time triggers03MCP serverModel Context ProtocolVisible SQL, audited logsBrazeEmail, SMSPush, in-appTriggered journeysClaudeEvery employeewith an AI seatasks in EnglishThird-party CDPRetired. $240K/yr saved.

What I built

  1. 01

    Warehouse

    Hourly BigQuery marts built from POS, loyalty (SessionM), app, web and Braze event streams. Identity keyed on mapped IDs rather than email strings. Version-controlled metric definitions in Git so 'active customer' means one thing everywhere.

  2. 02

    Segmentation layer

    A SQL-defined segment catalog that compiles to audiences and syncs to Braze on a schedule. Real-time triggers for the behaviors that matter (first order, lapsed, in-store scan). Email, SMS, push and in-app all read from the same segment definitions.

  3. 03

    MCP on top

    A Model Context Protocol server exposing the marts to Claude under least-privilege IAM. Every employee with an AI seat can ask questions in plain English and see the SQL that answered them. The daily query-log audit found no raw-dataset policy violations in the first months after rollout.

  4. 04

    Evaluation and quality

    A versioned evaluation set (49 cases: 29 rule, 6 ambiguity, 14 SQL) graded against cited rulings and reference SQL on a pinned window. Cases are retired, never edited, when a ruling changes, so expected answers never leak into the corpus. A backtest against a known-bad commit failed 4 of 6 affected cases, which is the proof it catches regressions. Behind it: a daily automated review of every query run against the warehouse (hundreds a day), a weekly drift detector for silent source-structure changes, and a deploy gate that blocks any view change whose live definition has drifted from the committed SQL.

Outcomes

  • $240K per year in CDP license fees eliminated. Pipelines now run on GCP at a fraction of that.
  • Identified-customer revenue up 17.7% year to date against 3.4% company growth, eight straight months of double-digit gains.
  • Identified share of sales up from 37% to 42%. Active customers (purchased in the trailing twelve months) up 14.6% to 550K, repeat orders up 19.4% on flat acquisition.
  • In-store loyalty scan program up 27% in orders. Mobile app adoption up 20%.
  • Governed data access went from a handful of SQL analysts to ten AI seats across the business.
  • No raw-dataset policy violations observed since the governance rollout.

Found along the way

  • Found and killed a silent race condition that had erased ten business days from a core fact table.
  • Corrected an eight-year-old employee-benefit classification error, surfacing roughly $750K in previously unflagged spend.

Stack

  • BigQuery
  • Cloud Run
  • Cloud Scheduler
  • IAM
  • Git
  • Braze
  • SessionM
  • Claude
  • MCP
  • Python
  • SQL
  • GA4 export
  • server-side GTM

Contact

If you need someone who can own the number and the stack, we should talk.

Open to: Director or VP of Data, Data Platform, or Analytics Engineering; VP or Senior Director of Marketing Technology; Head of Marketing Operations, Customer Data, or Growth Engineering; Advisory work on warehouses, identity, and AI-ready data.

brentlchristensen@gmail.comLinkedInGitHub801-699-5272Download resume (PDF)