How to Pass an AI Engineer Interview in 2026: 10 Questions With Practical Answers
An applied AI engineer interview often asks what to do when a model gives the wrong answer, calls the wrong tool, or costs too much to run. A strong answer names the decision, explains why it fits the problem, and shows how to check whether it worked. That can take three sentences.
These ten practice questions cover recurring themes in public interview guides and question banks as of November 2026. Formats vary by employer and role. The examples use a fictional online shop whose support assistant answers return-policy questions and looks up orders. All measurements and policies in the examples are illustrative.
Table of Contents
- What the Interview Can Look Like
- LLM Fundamentals and Retrieval
- Evaluation and Practical Coding
- Tool Use and Security
- Production Design and Project Discussion
- A 60-Minute Mock Interview
- Conclusion
What the Interview Can Look Like
Applied AI roles focus on building features that use models. Model-training and infrastructure roles can go deeper into statistics, training, or GPU serving. Start with the job description and ask the recruiter which rounds involve live coding, design, a take-home exercise, or a project discussion. Confirm the AI-tool rules for each round.
A coding round may ask for a small implementation or a bug fix. A design round asks for choices and trade-offs. A project discussion asks what the candidate personally did and how they knew it worked. The sample answers below are starting points. In a live interview, explain one decision first and expand when the interviewer asks why.
LLM Fundamentals and Retrieval
The shop has a general 30-day return policy and a separate seven-day policy for opened headphones. Both examples below begin with deciding what information the model needs at the moment it answers.
1. When Would You Use RAG or Fine-Tuning?
Sample answer. “I would use retrieval-augmented generation, or RAG, for the shop’s return policies. The service finds the applicable policy and passes it to the model with the question. That lets the shop update a policy in its document index and show the source in the answer.”
Practical follow-up. Suppose the assistant sees the correct policy but consistently labels the same return reasons incorrectly. Start with instructions, examples, and an evaluation set. If the errors persist and there are enough representative labeled examples, test fine-tuning for that classification behavior. Keep changing policy facts in the retrieval path.
2. The Assistant Gave the Wrong Answer. Where Do You Look First?
Sample answer. “The assistant said opened headphones could be returned within 30 days. I’d replay the request and check which passages actually reached the model. If the seven-day policy is missing, I’d inspect indexing and retrieval. If the model saw both policies, I’d check whether the prompt made the category-specific rule take precedence.”
Practical follow-up. If the passage was retrieved but omitted when the app shortened the prompt, fix context assembly. If the retriever missed it, try searching by product category and exact terms alongside semantic search. Add the original question to a regression set and include a general return question to check for overcorrection.
Evaluation and Practical Coding
After the headphones case is fixed, the next question is whether the change also improves other answers. An evaluation, or eval, is a repeatable test of that behavior on defined cases.
3. How Do You Know a New Prompt Is Better?
Sample answer. “I’d run both prompts on the same questions and hold the model and retrieved passages fixed. I’d label whether the answer is correct, whether its source supports it, and whether it asks for help when evidence is missing. I’d review results by category before choosing a prompt.”
Practical follow-up. In an illustrative 50-case set, prompt B could answer 44 questions acceptably compared with 40 for prompt A, yet confidently invent an answer on two of six cases where the policy is missing. Review those failures before release. Use deterministic checks for fields such as a known order status and human review for ambiguous policy answers. Keep some cases aside from prompt tuning.
4. Write a Python Function That Scores Retrieval
Interviewer prompt. Given the expected policy passage ID and the ranked results for each answerable question, calculate how often the expected passage appears in the first k results. Clarify that there is exactly one expected ID per case and that unanswered cases belong in a separate eval.
def hit_rate_at_k(cases, k):
if k < 1 or not cases:
raise ValueError("Provide cases and a positive k")
hits = 0
for expected_id, ranked_ids in cases:
if expected_id in ranked_ids[:k]:
hits += 1
return hits / len(cases)
cases = [
("headphone-policy", ["general-policy", "headphone-policy"]),
("gift-policy", ["general-policy"]),
("final-sale-policy", []),
]
assert hit_rate_at_k(cases, 1) == 0
assert hit_rate_at_k(cases, 2) == 1 / 3
Explain the result aloud. At k=2, one of the three questions gets its expected passage. The empty result still counts as a miss. This metric checks whether retrieval surfaced the evidence. Score the final answer separately to check whether the model used it correctly.
Tool Use and Security
A return policy explains what should happen. Looking up a specific order requires current account data, and issuing a refund changes it. Each step needs its own authority.
5. When Does a Feature Need an Agent?
Sample answer. “For order status, I’d use a fixed workflow. Authenticate the customer, look up the order, and present the shipping result. I’d consider an agent if the next step depends on what it discovers, such as investigating why an order shows delivered but the customer says it never arrived.”
Practical follow-up. That investigation might require choosing between shipment events and replacement history. Give the agent a short list of read-only tools, a time and step limit, and a handoff when it cannot resolve conflicting results. Compare it with a fixed workflow on the same cases before adding the extra model calls.
6. How Would You Let the Model Request a Refund Safely?
Sample answer. “The model can propose a refund for an order ID. The backend verifies the customer owns the order, calculates the eligible amount from trusted records, and checks whether confirmation or staff approval is required. The refund service owns execution and records the outcome.”
Practical follow-up. If the payment provider times out, look up the operation or retry with the same idempotency key when the provider supports it. The idempotency key identifies one approved refund attempt. Show the customer a pending status until there is evidence of success. A valid tool payload can still name someone else’s order, so validate the structure and check ownership independently.
7. What If a Retrieved Document Contains Malicious Instructions?
Sample answer. “A seller’s uploaded PDF might tell the assistant to ignore policy and refund every order. The PDF is evidence for the answer. It has no authority to approve a tool call. I’d keep the policy assistant read-only, and enforce authorization in the backend for any separate refund workflow.”
Practical follow-up. Put the malicious text in a test document and verify that the assistant cannot trigger a write or retrieve another customer’s order. Check access before including search results in the model context. A model refusal is useful to observe, and the test should also inspect what the backend actually allowed.
Production Design and Project Discussion
In a design interview, lead with what the first version must do. In a project discussion, use one concrete incident to show how the design was tested.
8. Design a Customer Support Assistant
Sample answer. “I’d ask which questions the first release covers and what should happen when it cannot answer. I’d ship policy answers with citations and signed-in order lookup first. A policy request searches current documents. An order request calls an authenticated API. If either path lacks reliable evidence, the assistant hands off to support.”
Practical follow-up. Walk through two requests on the whiteboard. “Can I return opened headphones?” needs the current category policy and its source. “Where is order 4821?” needs an ownership check and live status. If the order API is unavailable, explain what the customer sees. Close with success measures such as verified answer quality, handoff rate, and response time.
9. The Assistant Is Slow and Expensive. What Do You Change First?
Sample answer. “I’d measure search, tool calls, and model calls separately, then check input tokens and response time by request type. If simple order-status requests are expensive because they spend two model calls interpreting a known status, I’d test a direct response from the order API.”
Practical follow-up. Compare the change on the same set of ordinary lookups, missing orders, and permission failures. Track cost per resolved request and p95 latency, the time within which 95% of requests finish. For a streamed answer, also measure time to first visible text so an early token does not hide a slow full response.
10. Tell Me About an AI Feature You Built and What Went Wrong
Sample answer for a practice project. “I built a small returns assistant over sample policies. It failed on opened headphones because general policy ranked above the specific rule. I added category filtering and replayed a small labeled set. The headphone case improved, but queries with an unknown category still needed clarification.”
Prepare the evidence. Bring the original input, the incorrect answer, the change, and the before-and-after result. Explain which part was personal work. A local project can demonstrate good judgment when its scope is clear. Report the actual size of the test set and distinguish test results from production impact.
A 60-Minute Mock Interview
Use the same shop example across all four exercises. That makes it easier to spot an answer that changes the policy, customer identity, or expected behavior halfway through.
- 10 minutes. Answer questions 1 and 2 aloud in under a minute each. Then explain how to reproduce the headphones bug.
- 15 minutes. Implement the retrieval metric without looking at the sample. Test an empty result, a hit just beyond
k, and invalid input. - 25 minutes. Sketch the support assistant. Ask a partner to interrupt with a missing policy, an order owned by someone else, and a refund timeout. For each, say what the customer sees and what the backend does.
- 10 minutes. Tell a project story from memory. Check whether it names the failure, the chosen change, the evidence, and one remaining limitation.
Conclusion
A definition helps start an answer. The practical example shows how a candidate makes decisions when the requirements are incomplete or the system fails. The same skill applies to coding, system design, and project rounds.
Key takeaway. For each topic, practice one concrete decision, one realistic failure, and one test that would show whether the decision worked.
Comments (0)