We review portfolios with hiring partners every month. The same projects work and the same ones fail.
What works
- A messy public dataset cleaned in the open, with the decisions written down
- A small tool someone else uses weekly
- A rebuild of something the company you are applying to does, done badly on purpose, then fixed
What does not
- Titanic, Iris, or any tutorial dataset
- A dashboard with no question behind it
- A model with an accuracy number and no baseline
The test is simple: could a stranger tell, in thirty seconds, what decision your project helps someone make?