01 / The Goal
Start with what you're actually trying to do.
Before I build anything I want to know what the goal is. You can make a hundred dashboards, or train a model that scores beautifully, but if they're not pointed at what you're trying to decide, it's just noise. So I'd rather start there and work backwards to the data. That means pinning down the words too. What counts as a customer, when someone has actually churned. A model gets asked to predict exactly that, so it gets written down first.
02 / What it records
Every table is a record of a process, not of the world.
The usual starting position is a lot of data sitting in different systems and nobody fully trusting any of it. So I go and find what writes each row, when it fires, and the one everybody skips: what it does when nothing happens. A missing row is a claim too. And a model learns whatever the table says, not what actually happened. If a column only gets filled in after the outcome, the model is quietly reading the answer.
03 / Make it correct
Before analysis or a model, the raw data has to be right.
Duplicates, gaps, records that don't line up across systems. It isn't glamorous and it's most of the job, machine learning included: a model trained on bad rows is just wrong with confidence. What actually matters comes next: set up the checks so it stays clean, instead of me fixing the same thing every month. Checks that run on every load, and notice when new data stops looking like what the model learned from.
04 / The Model
Give it a shape, then teach it.
Two models, really. First the data model: facts in the middle, dimensions around them, one row meaning exactly one thing, so the same question stops getting two answers. Then, if the goal needs a prediction, the machine learning one. I start with the simplest thing that could work, hold back data it has never seen, and only reach for something bigger when it earns it on that test. Boring first. Boring is what somebody else can pick up.
05 / The so what
What sells the most isn't always what makes the most money.
Then the actual thinking. Ranking by revenue tells you what's popular; ranking by margin tells you what's worth pushing. Same with a model: a good score isn't the answer. What is it leaning on, and would you have it before the decision? A dip in the middle of the year is a question, not a finding. When the numbers are right people make better calls. When they're wrong, everyone's just guessing with confidence.
06 / The Dashboard
Open it, ten seconds, know where you stand.
It ends as something somebody actually uses. Usually a screen: whoever's running the business should understand its health in about ten seconds, then dig in. Sometimes it's a model on a production line, making the call in milliseconds. I build it backwards from the questions they'd actually ask, and keep watching it after it ships. When it stops answering the question, the goal at the top was wrong. Which is where you go back to.
