Covers the full ML maturity ladder — from notebook chaos to Level 2 automation — and how MLflow captures the reproducibility signals that make experiments auditable and repeatable. Tests whether you understand what can silently go wrong at each stage.

Explores how DVC and MLflow Model Registry work together to create an audit trail from raw data to promoted model. Traps include DVC garbage collection gotchas, registry stage semantics, and rollback vs. re-training distinctions.

Tests your ability to build lean, reproducible ML containers and wire them into a CI pipeline that actually catches model regressions — not just lint errors. Hard questions probe multi-stage builds, GIL-aware GPU CI queues, and training-serving skew detection.

From blue-green to canary to shadow — and from FastAPI to Triton. Covers the deployment lifecycle end to end including traffic splitting math, operating-threshold recalibration, GIL bottlenecks, and dynamic batching tuning. Designed to surface the gap between 'it works in staging' and 'it holds production SLAs'.

Digs into the operational realities of feature stores and pipeline orchestration — point-in-time correctness, online/offline skew, Airflow concurrency traps, and the pointer vs. XCom large-payload anti-pattern. Tests whether you can reason about data flow correctness, not just tool familiarity.

The hardest operational challenge in production ML: knowing when your model is wrong before your users do. Covers PSI, KS test, covariate vs. concept drift, multiple-testing problems in alert design, shadow mode blind spots, and business-metric vs. proxy-metric traps.

LLM-specific operational challenges: prompt versioning discipline, observability in RAG pipelines, token cost tracking, LLM testing pipelines, and deployment traps unique to generative models. Tests whether you understand why standard MLOps patterns need adaptation for LLMs.

A broad-coverage easy interview simulation. Tests your baseline fluency across the MLOps toolchain — from DVC checkout to blue-green rollback. Designed to feel like a 12-minute phone screen where the interviewer is checking whether you understand the fundamentals before going deeper.

Second easy mock. Focuses on the monitoring-to-LLMOps half of the syllabus — drift detection basics, model registry lifecycle, feature store online/offline concepts, and prompt versioning. Complements Set 1 for complete easy-tier coverage.

A mid-level interview simulation mixing applied reasoning, debugging scenarios, and architecture tradeoffs. Requires multi-step thinking — e.g., identifying why a CI gate never fails, what makes drift-triggered retraining loops dangerous, or why the previous Production model goes Archived not Deleted.

Second medium mock. Covers the operational and observability half — serving infrastructure tradeoffs, feature store skew, pipeline DAG design, monitoring alert design, and LLM cost/quality observability. Includes deceptive distractors that trap engineers who know the tool names but not the underlying mechanics.

A FAANG-level hard interview covering the full training-to-serving pipeline. Questions test edge cases in distributed training logging, DVC gc scope destruction, operating-threshold miscalibration after promotion, Python GIL serving bottlenecks, and Triton batching latency tuning. Expect scenario-based reasoning across infrastructure and ML simultaneously.

Second hard mock. Focuses on the post-deployment operational layer — feature store point-in-time violations, Airflow GPU pool starvation, PSI multiple-testing false-positive floods, KS effect-size vs. p-value traps, RAG component observability gaps, and prompt registry architecture. Senior-ML-engineer difficulty throughout.

Staff-engineer-level assessment across all 13 MLOps topics. Designed to distinguish senior engineers from staff/architect-level thinkers. Every question requires multi-step reasoning, understanding of failure modes under production load, and awareness of non-obvious system interactions. Covers: automated gate design flaws, platform primitive governance, compliance-grade data lineage, multi-GPU experiment logging, registry naming as interface contracts, non-root container security, tiered CI GPU queue management, counterfactual shadow-mode bias, GIL serving architecture, feature store consumer registry, Airflow idempotency under concurrency, importance-weighted drift alerting, business-metric vs. proxy-metric decoupling, and RAG component-level observability gaps.

The hardest assessment in the MLOps track. Every question is drawn from hard or high-medium difficulty and tests your ability to diagnose production failures — not describe tools. Scenarios include: automated evaluation gate that never fails (holdout leakage), GC destroying multi-branch DVC histories, canary evaluation window seasonality blindspot, Triton dynamic batching p99 tuning, point-in-time join violations causing silent 18% recall drops, PSI multiple-testing avalanche, KS statistical-vs-practical significance trap, RAG retrieval-quality monitoring gap, and LLM cost architecture. This test separates those who can talk about MLOps from those who can operate it.