Architecture, scalability, reliability, security, cost, and cross-team ownership.

A system-design answer

Say this first: AI vs ML vs deep learning is a choice between approaches with different strengths. The useful answer is the decision rule, not a dictionary definition.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply AI vs ML vs deep learning, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Choose the option that fits the workload and constraints; do not present one option as universally superior.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → AI vs ML vs deep learning → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: supervised vs unsupervised learning is a choice between approaches with different strengths. The useful answer is the decision rule, not a dictionary definition.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply supervised vs unsupervised learning, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Choose the option that fits the workload and constraints; do not present one option as universally superior.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → supervised vs unsupervised learning → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: classification vs regression is a choice between approaches with different strengths. The useful answer is the decision rule, not a dictionary definition.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply classification vs regression, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Choose the option that fits the workload and constraints; do not present one option as universally superior.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → classification vs regression → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: Overfitting memorizes quirks in training data and then fails on new data; underfitting is too simple to capture the useful signal.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply overfitting and underfitting, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → overfitting and underfitting → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: The important point about train validation test split is how an engineer recognizes the unsafe path early and prevents it from becoming customer impact.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply train validation test split, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Start with containment and evidence. Changing several variables at once makes the incident harder to understand.
• Call out prompt injection and unsupported answers and the control that reduces it.

Concrete check

npm test -- --runInBand

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → train validation test split → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: Precision asks whether positive predictions were correct; recall asks whether the system found the positives it should find. The right balance depends on the cost of each error.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply precision recall F1 and ROC-AUC, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → precision recall F1 and ROC-AUC → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: feature engineering and embeddings should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply feature engineering and embeddings, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → feature engineering and embeddings → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: tokenization and context window should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply tokenization and context window, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → tokenization and context window → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: prompt engineering basics should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply prompt engineering basics, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → prompt engineering basics → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: The important point about RAG pipeline fundamentals is how an engineer recognizes the unsafe path early and prevents it from becoming customer impact.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply RAG pipeline fundamentals, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Start with containment and evidence. Changing several variables at once makes the incident harder to understand.
• Call out prompt injection and unsupported answers and the control that reduces it.

Concrete check

SELECT COUNT(*) AS rows, MAX(loaded_at) AS freshest FROM <table>;

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → RAG pipeline fundamentals → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: Chunking splits source material into retrieval units. The boundary and size affect whether a retrieved passage contains enough context to support an answer.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply chunking strategy, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → chunking strategy → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: Similarity search represents content as vectors and retrieves nearby meanings; it is useful when users and documents do not use the same words.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply vector database and similarity search, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Concrete check

SELECT COUNT(*) AS rows, MAX(loaded_at) AS freshest FROM <table>;

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → vector database and similarity search → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: reranking and hybrid search should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply reranking and hybrid search, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → reranking and hybrid search → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: fine-tuning vs RAG is a choice between approaches with different strengths. The useful answer is the decision rule, not a dictionary definition.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply fine-tuning vs RAG, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Choose the option that fits the workload and constraints; do not present one option as universally superior.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → fine-tuning vs RAG → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: The important point about LLM hallucination control is how an engineer recognizes the unsafe path early and prevents it from becoming customer impact.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply LLM hallucination control, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Start with containment and evidence. Changing several variables at once makes the incident harder to understand.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → LLM hallucination control → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: prompt injection defense should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply prompt injection defense, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → prompt injection defense → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: batching and inference latency should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply batching and inference latency, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → batching and inference latency → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: model API integration should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply model API integration, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → model API integration → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: model evaluation framework should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply model evaluation framework, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → model evaluation framework → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: The important point about A/B testing AI features is how an engineer recognizes the unsafe path early and prevents it from becoming customer impact.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply A/B testing AI features, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Start with containment and evidence. Changing several variables at once makes the incident harder to understand.
• Call out prompt injection and unsupported answers and the control that reduces it.

Concrete check

npm test -- --runInBand

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → A/B testing AI features → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: model monitoring and drift should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply model monitoring and drift, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → model monitoring and drift → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: GPU serving cost optimization should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply GPU serving cost optimization, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → GPU serving cost optimization → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: privacy and PII handling should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply privacy and PII handling, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → privacy and PII handling → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: responsible AI and bias should be explained through its purpose, the boundary where it applies, and the evidence that shows it is working.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply responsible AI and bias, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• State the constraint that could change your decision, such as scale, data sensitivity, recovery target, or team ownership.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → responsible AI and bias → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.

A system-design answer

Say this first: The important point about agent workflow orchestration is how an engineer recognizes the unsafe path early and prevents it from becoming customer impact.

Use a real scenario

Imagine an internal support assistant that answers from approved policy documents. The team must decide how to apply agent workflow orchestration, verify the result, and explain the user impact. For an AI Engineer, attach the explanation to an evaluation set and retrieval trace.

Show judgment

• define boundaries, ownership, failure modes, and the operational feedback loop.
• Start with containment and evidence. Changing several variables at once makes the incident harder to understand.
• Call out prompt injection and unsupported answers and the control that reduces it.

Evidence to mention

Track grounded-answer rate and p95 response time. Say what baseline you compared against, what would trigger a rollback or escalation, and who owns the follow-up.

request or change → guardrail / validation → agent workflow orchestration → observable result → owner review

Practice prompt: Show how the team uses evaluation set and retrieval trace rather than relying on an informal agreement.