Three practices. Each one closes with a system.

01

Data engineering

When each report comes from a different place, the conversation goes to reconciling numbers. We set up the path of the data: from where it is born to the table the team opens. Orchestration runs in Airflow or Dagster, and the volume is processed in Databricks, BigQuery, or Spark, on the account you already have.

What gets built

• Ingestion and transformation with tests, not a loose script
• Orchestration in Airflow or Dagster, with retries and a place to see what failed
• Processing in Databricks, BigQuery, or Spark, without moving platforms for a demo

What stays. A pipeline that can be run again and a model another person can read.

02

AI applications

A chat stuck on the website is not a system. The agent or the chatbot does one concrete job — classify, draft, search your documents — and it is clear what a person reviews.

What gets built

• An agent with bounded tools and a defined input and output
• Evaluation with your examples, not a generic demo
• Limits: what the system must not answer, and where it stops

What stays. An application the team uses in its work, with a way to know if it gets worse.

03

Data science

We start from the decision, not the algorithm. If a careful count is enough, there is no model. If a prediction or an explanation is needed, the result can be defended in front of the person who decides.

What gets built

• The question and the metric, written before the analysis
• A reproducible analysis, with the data and the assumptions in view
• A recommendation of what to do, and what not to conclude

What stays. An answer that can be repeated when the data changes.