Blog

Hiring juniors

How to hire a junior AI engineer: the task, the marking and the interview

Most junior AI engineers will build with existing models, not train new ones, so quiz questions about model internals test the wrong thing. Here is a hiring process built around the real job: one task, five things to mark and one interview.

Wooble Team12 Nov 20268 min read

To hire a junior AI engineer well, test the job they will do: building something useful on top of existing AI models, with messy company material, and knowing when it is wrong. That takes one short task, a marking guide with five lines, and an interview about the candidate's own work. Quiz questions about model internals and timed puzzle rounds test a different job.

What does a junior AI engineer actually do?

In most companies, a junior AI engineer does not train models. They build with them. A typical first quarter looks like this:

  • Connect a model to the company's own documents so it can answer questions from them.
  • Write and adjust the instructions the model is given, and work out why an answer went wrong.
  • Build a way to check whether the answers are right, and keep checking as things change.
  • Handle the dull parts: documents in odd formats, old policies that contradict new ones, costs, slow responses.

The skill underneath all of it is judgement about a system that is right most of the time and confidently wrong the rest. That is what your process needs to find.

Why do the usual methods miss?

Theory quizzes reward people who memorised how the underlying models work. That is useful to know, but it is not the work, and it is the part an AI tool recites best.

Puzzle rounds test whether someone can solve a small algorithm problem under pressure. That says little about building something on messy documents.

The open take-home ("build a chatbot over these files") was the right idea. It now takes an AI tool a few seconds, so a working answer tells you nothing. We covered that in what to test now that AI can finish a take-home.

What task should you set?

Set a small version of the real thing, with traps in the material.

The brief: "Here is a set of our company documents. Build a small assistant that answers staff questions from them, and shows which document each answer came from. Then write a short note on what you did and where it falls short. Use any tools you like. It should take about an hour."

The material is where the test lives. Make sure it includes:

  • Two documents that disagree, such as an older policy and a newer memo that replaces part of it.
  • A question the documents cannot answer, so you can see whether the assistant admits it.
  • A document in an awkward format: a table, a scan, a long file with the important line buried in the middle.

Give each candidate a different set. Change the documents, the figures and where the traps sit. Then no answer can be shared, an AI tool cannot hand every applicant the same result, and you can still compare everyone because the task and the marking are the same.

How should you mark it?

Five things, each marked weak, acceptable or strong. Write them down before anyone applies.

What you markA strong answer
Does it work on the plain questions?Correct answers, each with its source
Does it handle the conflict?Uses the newer document and says why
Does it admit what it does not know?Says the documents do not cover it, and does not invent an answer
Did they check their own work?The note lists the questions they tried and what failed
Can they explain it?The note is clear about the approach and honest about its limits

Notice that only the first line is about the thing working. The other four are about judgement, which is the part that carries over to your codebase.

A candidate who used AI heavily and scores well on all five is a good hire. A candidate whose assistant works but whose note cannot say why is not.

What should you ask in the interview?

Skip the trivia and talk about their hand-in. Fifteen to twenty minutes is enough:

  • "Walk me through what you built, in the order you built it."
  • "These two documents disagree. What does your assistant do with that, and why?"
  • "What happens if I add a hundred more documents?"
  • "Show me a question it gets wrong."
  • "What would you do with another day?"

Someone who understood their work answers these comfortably. Our guide to five questions that show a candidate understood their own work goes through what good and poor answers sound like.

What should you not worry about?

  • Which tools they used. The field changes every few months. Judgement lasts longer than any one tool.
  • Their degree. Plenty of people doing this work well studied something else. HirePro's 2026 survey of Indian employers found 85% prefer graduates with AI skills, yet only 12% screen for them directly: most teams want the skill and have no way to see it. A task is a way to see it.
  • Polish. A plain assistant with an honest note beats a pretty one that makes things up.

The whole process on one page

  1. Post the job with the task as the way to apply.
  2. Each applicant gets their own documents and about an hour of work, any tools allowed.
  3. Mark all hand-ins against the same five lines.
  4. Interview from the top of the list, about each person's own work.
  5. Decide.

Wooble has this ready for the junior AI engineer role: the brief, each applicant's own documents and the marking are already written, you get a ranked list with the reason for every place, and the interview questions come from each person's hand-in. Your first hand-ins are free. See how it works.

See who can do the job, before you interview