An AI model at a client site suddenly loses performance. How would you reproduce the issue and determine whether the cause is the data, the model, or the deployment environment?

AI system debugging — isolating data drift, model degradation, and infrastructure issues in production ML systems.

Use a systematic isolation approach: (1) Check data pipeline freshness and schema changes. (2) Run the model on a known-good evaluation set to isolate model vs. data issues. (3) Compare inference latency and resource usage against baseline. (4) Check deployment config (quantization, batching, GPU memory) for environment drift.

• ✗Assuming the model is broken without checking data quality first
• ✗Not having a baseline evaluation set to compare against
• ✗Ignoring infrastructure metrics (OOM, thermal throttling, network latency)