Original practice scenario inspired by KLA Corporation's public business context; this is not a reported interview question or a description of its internal systems. A fictional electronics inspection model flags rare defects in images from manufacturing lots. Adjacent images from one lot look very similar. How would you assess model quality before proposing an inspection workflow?
Read model answer +
Evaluate at the operational decision grain and keep correlated lots separated. A high overall accuracy can conceal a model that misses almost every defect when negatives dominate.
Work through it
Split by manufacturing lot and time rather than randomly by image. Document labeling uncertainty and whether one defect can appear in several images.
Measure recall for defect categories, false alerts per inspected unit, and review workload across thresholds. Report uncertainty for rare categories with few positive examples.
Replay the proposed review workflow, including duplicate alerts and ambiguous images. Compare against the current inspection baseline and evaluate new equipment or process conditions separately.
Tradeoffs to discuss
More sensitivity usually increases review effort. Sampling extra defects helps diagnosis but requires prevalence-aware interpretation of precision.
Common pitfalls
Image-level leakage inflates generalization.
A balanced evaluation set cannot directly predict production alert volume without the real class prevalence.
Original practice scenario inspired by KLA Corporation's public business context; this is not a reported interview question or a description of its internal systems. A synthetic inspection pipeline produces large images and small defect metadata. A downstream archive slows during a production burst. Design buffering and prioritization without silently losing the link between images and results.
Read model answer +
Give each inspection a durable identity and track image and metadata persistence independently. Prioritization can reduce latency, but an incomplete record must remain visibly incomplete rather than appearing fully archived.
Work through it
Write metadata and image references to a durable manifest with checksums. Upload large blobs through a bounded queue, and record completion only after integrity verification.
Set queue limits and explicit overload policies. Preserve critical results while applying documented sampling or rejection rules to optional material, with counters for every loss.
Make consumers idempotent and reconcile manifests against blobs. Test sustained archive slowness, worker restarts, partial uploads, and recovery without duplicate records.
Tradeoffs to discuss
Local buffering buys time but requires capacity planning and crash recovery. Blocking production versus discarding optional data is a deliberate operating-policy choice.
Common pitfalls
Unbounded queues eventually exhaust memory.
A successful metadata write does not prove its referenced image exists or passed checksum validation.
