AI Platform and MLOps Engineer interview questions and answers

Sample answer

I start with the workload contract and representative evaluation set, then version the code, data references, model artifact, dependencies, and serving configuration together. The candidate must pass data and security checks, quality and safety thresholds, latency and cost limits, and a reproducibility check against the approved baseline. I promote an immutable candidate to a limited environment, observe it against production-relevant signals, and require an accountable approval before wider release.

Sample answer

I compare the offline and production request distributions, preprocessing, feature or retrieval versions, model configuration, and output handling. I inspect representative failures while respecting data access rules, then separate data or training-serving skew from serving defects and a weak evaluation set. I contain user impact if needed, add the missing cases to evaluation, and validate the correction on a shadow or limited rollout before replacing the current version.

Sample answer

Not by default. Drift is a signal to investigate, not proof that the model is harming outcomes or that fresh labels are available. I validate the signal, check input quality and business outcomes, compare performance by relevant segments, and determine whether the change is expected. Retraining creates a candidate that must pass the same evaluation, safety, and rollout gates as any other release; automation can prepare evidence without bypassing those controls.

Sample answer

A teammate should be able to identify the approved data snapshot or query, code revision, dependency and runtime versions, training parameters, random seeds where applicable, model initialization, and resulting artifact. I record lineage and evaluation outputs with the run, control access to sensitive data, and state which sources of nondeterminism remain. Reproducibility does not mean every accelerator run is bit-for-bit identical; it means the result and its evidence can be traced and meaningfully recreated.

Sample answer

I profile the complete request path first, including preprocessing, retrieval, queueing, model inference, postprocessing, and network calls. Then I test workload-appropriate options such as a smaller candidate model, optimized context or payload, batching, caching for valid repeated work, concurrency controls, or a different serving configuration. I compare task quality, tail latency, throughput, and cost per successful task on the same representative workload, and keep a safe fallback for overload.

Sample answer

I stop or reduce the rollout according to the release policy and assess the affected workflow and severity. I compare versioned metrics and representative examples for that segment, check whether traffic or evaluation coverage changed, and roll back or route to the last known-good version if that is the safest response. Before retrying, I add a segment-specific regression test and document the decision criteria so the aggregate score cannot hide the same failure again.