Case study

Deel: Rebuilding the Data Function at 90 → 2,000+ Employees

I joined Deel as its second data hire, a growth data analyst, while the company grew from 90 to 2,000+ employees. Analytics ran on a Postgres database and Metabase, every chart backed by its own hundreds of lines of hand-written SQL with no modelling layer underneath.

The transformation · Postgres to a source of truth

1. 01

   Arguing for a head of data

   I took the problem to leadership and argued for a dedicated head of data.

   Decision · Demand for answers grew with the company’s headcount, while chart-level SQL only grew with analyst hours. The fix had to start with how the function was staffed and led.

2. 02

   Stakeholder discovery

   I interviewed product, growth, and sales to define what each function needed from data.

   Decision · The existing charts recorded what had once been asked for; requirements had to come from the business, not from what Metabase already showed.

3. 03

   Re-platforming

   SnowflakeFivetran

   With the new data leader on board, we moved warehousing onto Snowflake and put ingestion on Fivetran.

   Decision · Analytics needed its own warehouse and managed ingestion, so the effort could go into modelling instead of moving data.

4. 04

   Modelling the source of truth

   dbtSnowflake

   I dimensionally modelled the warehouse in dbt: version-controlled, tested, one definition per metric.

   Decision · Every number defined once in the model, instead of once per chart in hand-maintained SQL.

5. 05

   Self-service BI

   Looker

   We rebuilt reporting in Looker on top of the modelled layer; teams answered their own questions, and ad-hoc analysis requests fell by roughly 60%.

   Decision · Once definitions live in the model, BI becomes an access layer: the same numbers for everyone, without an analyst in the loop.

What shipped

• Snowflake warehouse with Fivetran ingestion, replacing analytics run directly against Postgres
• A dimensionally modelled dbt project: tested, version-controlled, a single definition per metric
• Looker self-service BI built on the model, replacing per-chart SQL in Metabase
• Documented data requirements for product, growth and sales, from direct stakeholder discovery
• A data function under dedicated leadership, from a hire I made the case for

Stack

Reporting went from hundreds of lines of SQL per chart to one definition per metric in the model, and ad-hoc analysis requests fell by roughly 60%.