Business Context

NovaCircuits manufactures industrial control boards across 6 production lines and wants a simple, explainable model that predicts whether a board will fail final quality inspection. The operations team does not just want predictions—they want an engineer who can clearly explain the mechanics of the model and how process variables influence defect risk.

Dataset

You are given a historical dataset of board-level production records collected over 18 months.

• Rows: 92,000 manufactured boards
• Target: defect_flag — whether the board failed final inspection
• Class balance: 11.4% defective, 88.6% non-defective
• Missing data: ~7% missing in humidity and some sensor channels due to intermittent logging failures; <2% missing in operator fields

Success Criteria

A good solution should achieve strong ranking and classification performance while remaining interpretable enough for process engineers to act on. Target at least ROC-AUC > 0.84, F1 > 0.55, and recall > 0.70 for the defective class after threshold tuning.

Constraints

• Predictions must be explainable to non-ML manufacturing engineers
• Batch scoring for ~8,000 boards/day; latency is not strict, but reliability is
• The model should be easy to retrain weekly and robust to moderate feature drift

Deliverables

1. Build a baseline defect-classification model using an interpretable supervised learning approach.
2. Explain the underlying mechanics of the model, including how features affect predictions.
3. Describe preprocessing for missing values, categorical variables, and scaling.
4. Evaluate performance with appropriate metrics for an imbalanced binary classification task.
5. Summarize key tradeoffs between interpretability, calibration, and predictive power.