AI / ML Interview Questions

MLOps applies DevOps principles to the ML lifecycle: data versioning, experiment tracking, model registry, continuous training, deployment, and drift monitoring. DevOps focuses on app code and infra; MLOps also manages data, models, and evaluation metrics that change over time.

Typical stages: data ingestion and validation → feature engineering → training/experiment tracking → model evaluation → registry → CI/CD packaging → deployment (online/batch) → monitoring (data/concept drift) → automated retraining triggers.

Track input distribution shifts (data drift), prediction/label shifts (concept drift), and performance metrics. Use statistical tests or PSI/KS, set alert thresholds, then retrain, fall back to a previous model, or add human review depending on severity.

A model registry stores versioned models with metadata (metrics, data lineage, stage: Staging/Production). It enables reproducible promotion, rollback, auditability, and team collaboration — similar to an artifact registry for ML models.

Online (real-time API) for low-latency user-facing decisions. Batch for nightly scoring, recommendations dumps, or large offline jobs. Many systems mix both: batch features + online serving.

Separate data/model pipelines from app CI. Include unit tests, data schema checks, training reproducibility, offline eval gates, container build, canary/shadow deploy, and automated rollback on metric regression.

LLMOps is the practice of deploying and operating large language model applications in production: prompt/version management, RAG pipelines, fine-tuning, evaluation, latency/cost monitoring, safety guardrails, and continuous improvement.

Ingest documents → chunk → embed → store in vector DB → retrieve top-k (often hybrid search) → optional re-rank → inject context into prompt → generate with citations. Add evals for retrieval quality and answer faithfulness.

Use RAG for up-to-date private knowledge and citations. Fine-tune for style, domain language, structured behavior, or tool-use patterns. Many production systems combine light fine-tuning with strong RAG.

Use offline eval suites (golden Q&A), retrieval metrics (recall@k), answer faithfulness/groundedness, latency and cost budgets, plus online feedback (thumbs, task success). Prefer automated graders + spot human review.

Cache frequent prompts, use smaller/faster models for routing, compress context, quantize local models, batch where possible, set max tokens, and use retrieval to avoid stuffing huge contexts.

Input/output filtering, PII redaction, allowlisted tools, rate limits, human-in-the-loop for destructive actions, audit logs, RBAC, and policy checks before external API calls.

AIOps uses AI/ML on telemetry (metrics, logs, traces, events) to detect anomalies, correlate incidents, predict failures, and automate remediation — reducing alert noise and MTTR.

Baselines from historical seasonality, statistical thresholds, or ML models (isolation forest, forecasting residuals). Alerts fire on deviations; multi-signal correlation reduces false positives.

Correlate alerts across services using topology/time windows, cluster related events, rank likely causes, and suggest runbooks. Graph/dependency context improves accuracy over single-metric alerts.

Monitoring checks known conditions (CPU > 90%). Observability lets you ask new questions using metrics, logs, and traces together — essential for complex distributed/ML systems.

Detect signal → classify severity → run playbook (restart pod, scale, rollback) with approval gates for high risk → notify Slack/Jira → log outcome for learning. Always include kill switches.

A chatbot mainly generates text. An agent plans, uses tools/APIs, maintains state/memory, and completes multi-step goals. Agents need orchestration, permissions, and evaluation beyond chat quality.

The model reasons, selects a tool, observes the result, then continues until a stop condition. Production agents add retries, timeouts, tool allowlists, and structured schemas for tool args.

MCP is a standard way for models/agents to connect to tools and data sources (servers/clients). It helps enterprises plug agents into GitHub, Slack, DBs, and internal APIs consistently.

Short-term: conversation/state in the session. Long-term: vector store or knowledge graph for facts/preferences. Persist tool results carefully, avoid leaking secrets, and expire stale memory.

Unit-test tools, simulation evals for workflows, adversarial prompts, latency/cost budgets, and human review for critical paths. Measure task success rate, not just fluency.

An FDE sits close to customers: scopes use cases, integrates data sources, customizes agents/pipelines, ships production pilots quickly, and feeds product feedback. Strong systems + customer skills matter.

Define schemas/SLAs, idempotent jobs, validation checks, lineage, retries/DLQs, monitoring, and backfills. Prefer incremental processing and clear ownership of source → feature store → model.

Batch is simpler and cheaper for periodic features. Streaming enables near-real-time signals but needs state, late data handling, and higher ops cost. Pick based on freshness requirements.

Harden auth/secrets, add observability, define SLOs, automate deploy, write runbooks, add eval/regression tests, document handoff, and plan capacity/cost. Convert notebooks into services/pipelines.

High bias (underfit): richer features, more complex model, longer training. High variance (overfit): more data, regularization, dropout, simpler model, cross-validation.

Precision: minimize false positives (spam filter). Recall: minimize false negatives (fraud/cancer detection). Use F1 when both matter; choose thresholds from business cost.

Embeddings map text/images into vectors capturing semantic similarity. Used for search, RAG, clustering, recommendations, and anomaly detection.

Transformers use self-attention to weigh token relationships, enabling parallel training and strong language modeling. Core stack: tokenization → embeddings → attention blocks → output head.

Strong retrieval, citation requirements, refusal when context is weak, grounded prompts, faithfulness evals, and temperature/control settings. Never claim unsupported facts.

Cover candidate generation, ranking model, features (batch + realtime), A/B testing, feedback loop, caching, monitoring CTR/latency, and fallback popularity models.

Connectors → ACL-aware indexing → chunk/embed → hybrid retrieval → re-rank → answer with citations → feedback → admin console. Enforce permissions at retrieval time.

GPU node pools, job queue (Kubeflow/Ray), experiment tracking, artifact store, secrets, network policies, autoscaling, cost attribution, and notebook → job conversion paths.

Multi-replica inference (vLLM), load balancer, health checks, autoscaling on queue depth, model warm pools, canary rollout, circuit breakers, and regional failover if needed.