Vol. 01 · 2026
Data engineer and analyst
Most of the damage in an analysis is a definition that slipped. I want the warehouse and the readout to be the same person’s problem: choose the grain, name the metric, and sit in the room when someone challenges the number.
Turn the card.
Log
Unfold a margin.
I run the mart and I still write the SQL when the question is sharp. Other people add models. I review the ones that would change a number the board has already learned.
A weekday refresh with a named owner, and a dictionary short enough to read.
Half the job was the pipeline. The other half was sitting with the people who answered the phones when a provider called about a claim. The claims cohort study started because those calls and the dashboard disagreed.
A tail metric the operations review still uses.
I learned to show the error before the chart. Merchants did not want a model that embarrassed them in front of their director. They wanted one that would admit, in writing, where it was worse than their spreadsheet.
A hybrid forecast that let staples be modeled and promotions stay human.
Instruments
Observations
If I cannot regenerate the figure from the warehouse, I do not take it into a meeting. Screenshots travel faster than definitions, and they do not age honestly.
Get the grain of the table wrong and every average lies politely. I would rather spend the first week on the grain than the last night on the color of a bar.