Google Machine Learning Engineer Interview Questions

I would assume that both user input and retrieved content may be adversarial. My design would separate trusted instructions from untrusted content and apply strict boundaries when passing information to the model. The reason I choose this approach is that prompt injection is often a trust-boundary problem, not just a model problem. I would also think about tool access control, content filtering, and logging suspicious prompts for review. If the interviewer wants more depth, I would mention red-team testing.

I would compare the offline data distribution with real production traffic to see whether the evaluation set is representative. Then I would inspect feature availability, preprocessing differences, and feedback loops that may not exist in the offline test. The reason I choose this approach is that a model can look strong offline while failing on the cases that matter in production. I would also think about delayed labels and segment-level performance. If needed, I would mention shadow deployment or canary rollout.

I would monitor input drift, output quality proxies, latency, errors, and user feedback. I would also build segment-level dashboards so I can see whether the problem is limited to one slice of traffic. The reason I choose segment-level monitoring is that aggregate metrics often hide serious regressions. I would think about alerting thresholds, periodic human review, and comparing canary traffic against a stable baseline. If the interviewer wants more detail, I would mention version tracking for data, features, and model weights.

I would improve retrieval quality first, because weak retrieval often causes the model to guess. Then I would narrow the answer scope, require evidence, and add refusal behavior when support is missing. The reason I choose this approach is that hallucinations usually happen when the system asks the model to invent facts instead of grounding the answer. I would also think about post-generation validation and confidence-based fallbacks. If the interviewer wants more depth, I would discuss evaluation on factual consistency.

I would check whether the data, preprocessing, hyperparameters, or environment changed from the baseline. Then I would compare the failed run against a known good run and look for differences in seeds, shuffling, optimizer settings, or input handling. The reason I choose this approach is that most regressions come from a small number of change points. I would also verify the evaluation script, because sometimes the model is fine and the metric is wrong. If needed, I would explain how I would make the experiment more reproducible.