Data scientists are fundamentally problem solvers.

The best way to assess technical problem solving is a structured hackathon. That is, to be given

* a problem with multiple subproblems and solution milestones

* with both objective and subjective criteria

* freedom of tools

* a "junkyard" of resources

* a fixed amount of time for each deliverable of the problem

And then you observe the process and the results.

For data science, the subproblems should be:

* requirements gathering / understanding the problem

* data acquisition, prep, and analysis

* refinement of requirements / communication

* feature engineering

* modeling

* presentation / storytelling / viz