How to Hire an AI/ML Engineer in Dubai in 2026

Q1: How would you build a production RAG system handling 10,000 queries per day?

What to assess: Strong answer: chunking strategy (semantic vs. fixed-size), embedding model choice with justification, vector store selection, re-ranking layer, caching for repeated queries, monitoring for hallucinations. Red flag: candidate describes only the happy path.

Q2: Your fine-tuned model degrades in production after 2 weeks — what do you do?

What to assess: Look for: data drift analysis, model monitoring setup, evaluation pipeline for regression detection, and a plan for continuous fine-tuning or RAG augmentation. Senior engineers mention shadow deployment and A/B testing.

Q3: Explain the difference between fine-tuning, RAG, and prompt engineering — when to use each.

What to assess: Prompt engineering: fastest, cheapest, limited by context. RAG: retrieves fresh knowledge at inference time. Fine-tuning: changes weights for consistent style or domain reasoning. Strong candidate gives concrete use cases for each.

Q4: How do you evaluate an LLM application when ground truth is expensive to obtain?

What to assess: Expect: LLM-as-judge pipelines, RAGAS for RAG evaluation, human-in-the-loop sampling, embedding similarity metrics, business proxy metrics. Red flag: only mentions BLEU/ROUGE.

Q5: Reduce LLM API latency from 8 seconds to under 2 seconds — walk me through your approach.

What to assess: Good answers: streaming responses, prompt compression, smaller model for simple queries, request caching, async pre-generation. Principal engineers also discuss distillation and quantisation.