Google Machine Learning Engineer Interview Questions

I would use ranking metrics such as NDCG, MAP, recall@K, precision@K, and possibly calibration or coverage depending on the product. The reason I use more than one metric is that ranking quality has multiple dimensions, and one score rarely tells the whole story. I would also think about business metrics such as click-through rate, conversions, or retention if the model is tied to a product surface. In an interview, I would explain that offline ranking metrics should be paired with online experimentation whenever possible. I would also mention that the chosen metric should match the user experience we are trying to improve.

I would separate offline feature computation from online serving while keeping the definitions consistent. The store would need a way to version features, validate schemas, and serve low-latency lookups for inference while also supporting batch access for training. The reason I choose this design is that training-serving skew is one of the biggest causes of ML bugs in production. I would also think about freshness, backfills, and feature lineage so we can understand where each value came from. If the interviewer wants more depth, I would discuss point-in-time correctness and how to handle missing features.

I would first profile the pipeline to see whether the bottleneck is data loading, preprocessing, model compute, or hardware utilization. If the model is underusing the accelerator, I would look at batch size, mixed precision, and input pipeline efficiency. The reason I choose measurement first is that training slowdowns often come from the data path rather than the model itself. I would also think about distributed training, gradient accumulation, and whether the architecture itself is too expensive for the target latency or training budget. If needed, I would compare speed gains against possible quality loss.

I would first determine whether the model is hallucinating because it lacks knowledge, lacks grounding, or is being prompted poorly. A strong fix is to improve retrieval, narrow the scope of the answer, and require citations or evidence where appropriate. The reason I choose this direction is that LLMs tend to invent details when they are forced to guess. I would also think about confidence thresholds, refusal behavior, and post-generation validation. If the interviewer wants more depth, I would mention evaluation sets focused on factual consistency.

I would monitor the input distribution, output distribution, latency, and downstream business metrics so I can detect both technical and product regressions. I would also break metrics down by segment because aggregate averages can hide important failures. The reason I choose this monitoring approach is that drift is usually gradual and easy to miss if you only look at final outcomes. I would think about alerting, canary comparisons, and retraining triggers. If the interviewer wants more detail, I would explain how I would distinguish data drift from concept drift.