Design model deployment, monitoring, and low-latency inference
Company: Capital One
Role: Machine Learning Engineer
Category: ML System Design
Difficulty: medium
Interview Round: Onsite
You have trained a fraud detection model and need to productionize it.
## Part A: Deployment
- How would you deploy an ML model to production?
- What artifacts do you version and how do you enable safe rollouts/rollbacks?
## Part B: Monitoring
- After deployment, how do you monitor the model?
- What metrics do you track for:
- service health,
- data quality/drift,
- model performance,
- business impact?
## Part C: Latency SLO
The model is deployed behind an online API, but it is missing a strict latency requirement: **p99 latency < 50 ms**.
- How do you diagnose where time is spent?
- What concrete changes would you consider across features, model, infrastructure, and serving to meet the SLO without unacceptable accuracy loss?
Quick Answer: This question evaluates competency in ML system design and production engineering—covering model deployment and versioning, safe rollouts and rollbacks, monitoring of service health, data quality/drift, model performance and business metrics, and latency optimization for low-latency inference.
