AI Assessment and the Review Interview

What they send

After the first technical, an AI assessment arrives by email. For the NLP roles it has been three separate tasks in one brief, usually a supervised text task on a dataset they provide, a retrieval or summarisation component, and a small write up of your approach. You get days, not hours, and you submit a notebook or a repository.

The interview that follows

This is the part people are not ready for. The next technical has one agenda: your submission. Expect, for every choice you made:

• Why this model and not a simpler one? What was your baseline, and did you beat it?
• What other approaches could have solved this task, and what is the difference between them? Name two or three alternatives with their trade offs, not one.
• Why this metric? What does it hide?
• How did you split the data, and are you sure nothing leaked?
• What would you change with one more week, and what would you change to put this in production?

They use the same pattern on your CV. Every project you list gets the same treatment: which model, why that one, do you understand how it works internally.

Examples

Example 1 (strong) A short README listing the three approaches considered, a table of metrics per approach, one paragraph on why the chosen one won, and a section titled "what I would do next". The interview walked through that file and finished early.

Example 2 (weak) A notebook with the best score and no history. Asked why not a classical baseline, the candidate had not tried one and could not say whether the transformer was earning its cost.

Tips

• Keep every failed experiment in the repository or at least in the README. The alternatives are the interview.
• State the assumptions you made about the ambiguous parts of the brief instead of quietly picking one.
• Leave the code runnable end to end. They open it with you on the call.
• Budget for the wait: the full process has run from one month to four, and the security clearance sits at the end of it.