The problem
Alma is a French buy-now-pay-later fintech. Its dashboards ran straight on the application database and raw data: hard to maintain and slow for analysts to work with. Data scientists lacked the tooling and data to industrialize their risk models, and business teams needed data pushed into the tools they used every day. Its only data engineer was leaving, and he couldn’t both develop its data platform and handle its operational data needs.
What I built
As one of two lead data engineers, I built Alma’s complete modern data platform on GCP: raw data pulled from every relevant source daily or hourly, analytics tooling, and layered dbt models up to business-ready marts, with the stakeholder dashboards built on top.
Infrastructure. I set up, in Terraform, every infrastructure asset a reliable data platform needs: Airflow, Kubernetes, Airbyte, BigQuery, dbt, and change data capture (CDC) on Postgres.
Ingestion. I brought six sources into the warehouse: the Postgres application database, HubSpot, Zendesk, Stripe, Adyen and Google Sheets. The app database started as daily batch loads and moved to change data capture once teams needed fresh data, cutting its freshness from 24 hours to 5 minutes. Data-quality dashboards compared the warehouse with each source of truth, and I monitored the pipelines daily.
Data back into the business tools. Custom reverse-ETL connectors sent modeled data to HubSpot, Zendesk and Customer.io, and third parties got their own GCP project to deliver data without touching production.
Access control. A data leak at a payments company exposes merchants and their customers at once. Tables and columns were tagged by sensitivity in BigQuery, so only accredited people could read sensitive data.
A modeling layer analysts can own. With the business analysts and stakeholders, we built layered dbt transformations, from raw data up to business objects, then moved the existing dashboards onto them and trained analysts and data scientists to build their own models without a data engineer.
Changes that can’t break production silently. CI ran statistical checks on every pull request: if a column’s distribution shifted, the merge was blocked. It also started Airflow and dbt in Docker Compose and blocked the merge on any failure.
Fraud-detection model. I also coached the data science team on engineering best practices, and we industrialized their risk-assessment model together: the one that decides whether a transaction is accepted. An A/B test showed the new version avoids about $50M of fraudulent transactions a year, a result that would not have been possible without the platform underneath it.
Who runs it now
When Alma hired a data engineer four months in, I coached them and we built the rest of the platform together. They became the lead data engineer, Alma hired more people around them, and I handed the platform over to that internal team.


