Design an OOD detection system
Overview
This question evaluates a candidate's competency in ML system design, specifically out-of-distribution detection, production monitoring, interpretability, and feedback loops for maintaining model reliability and data management.
Prompt
You are building a product that uses an ML classifier in production (e.g., for routing, ranking, safety, fraud, or categorization). Over time, the live input distribution may shift and users may submit inputs that are out-of-distribution (OOD) relative to the model’s training data.
Design an end-to-end system to identify OOD data in production and support actions such as alerting, safe fallback behavior, and data collection for retraining.
Requirements
- Detect OOD inputs in (near) real time and/or via batch monitoring.
- Minimize false alarms while still catching meaningful distribution shift.
- Provide interpretable signals to on-call/ML engineers (what changed, where, and how severe).
- Support a feedback loop: triage → labeling (if needed) → retraining/evaluation.
What to cover
- Define what “OOD” means for this product (vs. mislabeled, rare-but-in-distribution, adversarial, or novel classes).
- Propose modeling/algorithmic approaches for OOD detection.
- Specify offline evaluation and online metrics.
- Design the data/serving/monitoring architecture.
- Decide what happens when an input is flagged OOD (fallbacks, user experience, logging).
- Handle edge cases: class imbalance, seasonality, new features, model updates, and cold start.
Assume you can log inputs/embeddings/predictions and you have a standard feature store + model serving stack.
Solution
Login required