Interview prep · 🤖 Data & AI Engineer
Data & AI Engineer interview questionsand answers (30, by level).
The questions Indian employers actually ask for this role — fresher, mid-level and senior — with model answers written by engineers who sit on the other side of the table. Use with the role path and the 30-day lab plan.
Fresher / L1
Fundamentals they test on day one
1ETL vs ELT?
ETL transforms before loading (legacy warehouses); ELT loads raw then transforms in the warehouse (modern, dbt).
2What is a data warehouse vs data lake?
Warehouse: structured, modelled for analytics (BigQuery, Snowflake); lake: raw files at scale (S3/Parquet); lakehouse merges them.
3Explain star schema.
Fact table (measures) linked to dimension tables (who/what/when); fast, understandable analytics.
4Pandas vs SQL for transformation?
SQL for set-based work in the DB; pandas for in-memory manipulation; push heavy work to the engine.
5What is an embedding?
A vector representing meaning of text/image; similar items are close; used for search and RAG.
6Supervised vs unsupervised learning?
Supervised learns from labelled data (classification/regression); unsupervised finds structure (clustering).
7What is overfitting?
Model memorises training data, fails on new data; fix with more data, regularisation, cross-validation.
8Parquet vs CSV?
Parquet: columnar, compressed, typed — far faster for analytics; CSV: universal but slow and untyped.
9What is Airflow?
Workflow orchestrator defining DAGs of tasks with scheduling, retries and monitoring.
10Precision vs recall?
Precision: of predicted positives, how many correct; recall: of actual positives, how many found; F1 balances.
2–5 years / L2
What a working engineer must know
1Design a pipeline from 20 sources to a warehouse with daily SLAs.
Ingestion (Airbyte/Fivetran or custom), landing in object storage, dbt models with tests, orchestration (Airflow/Dagster), data quality checks, lineage, alerts on SLA.
2How does RAG work and what makes it good?
Chunk → embed → vector index → retrieve top-k → prompt LLM with context; quality depends on chunking, retrieval evaluation, reranking, citations.
3Choosing an LLM: hosted vs open-weights?
Cost, latency, data residency, capability; open models (Llama/Qwen) on-prem for sensitive data; see our LLM guide.
4Data quality framework?
Expectations (Great Expectations/dbt tests): freshness, volume, schema, nulls, uniqueness; quarantine bad data; dashboards.
5Spark vs DuckDB vs warehouse SQL?
Spark for distributed big data; DuckDB for fast single-node analytics; warehouse SQL for shared governed analytics.
6Feature store — what problem does it solve?
Consistent features between training and serving, reuse, point-in-time correctness.
7How do you evaluate an LLM app?
Golden datasets, automated judges plus human review, metrics (faithfulness, relevance), regression tests per prompt change.
8Streaming vs batch — when?
Streaming (Kafka/Flink) for low-latency needs (fraud, alerts); batch for cost-efficient periodic analytics.
9Vector database choice?
pgvector for most; Milvus/Qdrant/Weaviate for very large scale or special filters; consider ops burden.
10Privacy in data pipelines (DPDP)?
Minimise, pseudonymise/mask, access controls, retention, consent flags propagated, audit.
Senior / L3
Design, judgement and leadership
1Design an AI agent for IT operations with safety.
Tool-calling LLM over monitoring/ticket APIs, read-only by default, approval gates for actions, audit log, evals, fallbacks; see how we run ours.
2MLOps: how do you take a model to production and keep it there?
Versioned data/models, CI for training, registry, canary deploy, monitoring drift/latency, retraining triggers, rollback.
3Data platform governance across departments?
Catalog, ownership, access policies, lineage, SLAs, cost allocation, data contracts.
4GPU infrastructure on-prem vs cloud for an Indian company?
On-prem for steady training/inference (Proxmox + passthrough, India power costs), cloud for bursts; consider IndiaAI compute; model sizing vs VRAM.
5Cost control for LLM products?
Caching, smaller models for routing, prompt compression, batching, token budgets, monitoring per feature.
6How do you prevent prompt injection in a RAG app?
Treat retrieved content as data, system prompt hardening, output validation, tool permission scoping, red-team tests.
7Real-time analytics architecture for 1M events/min?
Kafka → stream processor (Flink) → OLAP store (ClickHouse/Druid) → dashboards; exactly-once semantics, backpressure.
8Migrating a legacy reporting stack to a modern one?
Inventory reports, define metrics layer, migrate sources, parallel validation, train users, retire.
9Responsible AI — what do you implement?
Bias checks, explainability where required, human oversight, data consent, model cards, incident process.
10How do you measure the business value of a data team?
Decisions enabled, time saved, revenue/cost impact per product, adoption of datasets, SLA adherence.
How to answer
Four habits that pass interviews
- Say the structure first“I’d check three things: …” — interviewers reward a method more than a lucky answer.
- Numbers and namesMention real tools, versions and figures from your lab; vague answers lose.
- Admit gaps cleanly“I haven’t run X in production; here is how I’d approach it” beats bluffing.
- Ask about the environmentScale, on-call, tooling — it shows you think like an operator.
Ready to be hired, or ready to hire?
Engineers: join our vetted network for projects and placements. Companies: we shortlist data & ai engineers within 48 hours.