AI engineer assessment by conversation, not multiple choice

What the assessment covers

The ground an AI engineer is expected to hold, taken from the five roles the diagnostic carries questions for: ML Engineer, LLM Engineer, Applied AI Engineer, AI Infrastructure Engineer, and MLOps. General software engineering sits outside it, and so does anything the question bank cannot support — an assessment that stretches past its evidence is guessing with a confident tone.

• Model training, evaluation design, and choosing the metric that matches the decision
• LLM application ground: retrieval, grounding, and how those fail in production
• Serving, latency, and cost trade-offs under real traffic
• Monitoring, drift, and triaging a system that worked last quarter

What it opens with, and why there is no question list

It starts somewhere ordinary for your role — how you would evaluate a retrieval step on its own, why a model that scored well offline degraded after release, what you would check first when serving latency doubled. Those are openings, not the assessment. What is being measured is the second and third question, which are written from your answer and therefore cannot be printed in advance.

Explanation as the measurement

You are asked to explain a concept in your own words. There are no options to select between, nothing to eliminate, and no way to pattern-match to a familiar-looking answer. If the explanation thins out under a follow-up, that is the signal.

• No answer options to recognise or eliminate
• Follow-ups probe the specific claim you just made
• Depth is assessed categorically, never as a percentage

What the fifteen minutes look like

You pick a role, you talk, and Veda follows. Clean explanations are marked and left behind; a fluent but hollow one gets pressed until it either holds or gives way. There is no timer to beat and no counter telling you how far through you are, because the path length depends on where your answers lead. When the session closes, the map is waiting.

Honest coverage reporting

The result separates three states that a single score would blur together: concepts you explained clearly, concepts where the explanation ran out, and concepts the session never reached. The third category is reported as untested rather than being folded into an aggregate that would misrepresent it.

Where this parts company with a scored screening test

A screening test exists to rank one candidate against others for someone who is hiring, so it needs a fixed instrument and a comparable number at the end. This has neither, and the absence is deliberate rather than a missing feature.

• Nothing is sent to an employer, because no employer is on the other end
• No cut-off, percentile, or comparison against other engineers
• Every gap it names shows the reasoning behind it, and you can dispute it

Who should skip it

If you are hiring and need a scored instrument to compare applicants, this is the wrong tool and will frustrate you within a minute. Same if you want a bank of questions to memorise, or if your interviews are general backend, DSA, or behavioural. It is worth your time when you have prepared for AI/ML rounds already and cannot tell which part of that preparation is thinner than it feels.