Wingman MLOps Engineer Interview: Rounds, Questions and Preparation

What the job involves

As an MLOps Engineer at a fast-growing AI or deeptech company, you will build and maintain the infrastructure that takes machine learning models from experimentation to reliable production. Day to day, you will automate training pipelines, manage model deployments, monitor model performance and drift, and work closely with data scientists and backend engineers to ship AI features. Expect to spend time writing code for CI/CD, container orchestration, cloud services, and observability.

Role and Technical Knowledge

1. 1

   What is the difference between DevOps and MLOps, and why does MLOps need additional tools?

   A strong answer covers that MLOps extends DevOps to handle data versioning, model training, model registry, and monitoring for drift, not just code deployment. Mention tools like MLflow or DVC for tracking experiments and data.

2. 2

   How would you containerize a simple machine learning model and serve it as an API?

   Explain writing a Flask or FastAPI app that loads the model, creating a Dockerfile with the right base image, exposing a port, and testing locally with curl. Mention handling dependencies and keeping the image small.

3. 3

   What is model drift, and how can you detect it in production?

   Define data drift and concept drift. Describe monitoring input feature distributions and prediction distributions over time, using statistical tests or tools like Evidently AI, and setting alerts when drift exceeds a threshold.

4. 4

   Explain how you would set up a CI/CD pipeline for a machine learning project.

   Cover stages: code linting and unit tests, data validation, model training, evaluation against a baseline, and deployment to staging or production. Mention using GitHub Actions or Jenkins and storing artifacts in a model registry.

Practical Scenarios

1. 5

   Your model's accuracy dropped suddenly in production. How do you debug and fix it?

   Start by checking data quality and schema changes, then compare recent input distributions to training data. Look at logs for errors, verify the model version deployed, and if drift is confirmed, retrain with recent data and roll back if needed.

2. 6

   You need to deploy a model that must handle 1000 requests per second with low latency. What architecture would you propose?

   Suggest containerizing the model, deploying on Kubernetes with horizontal pod autoscaling, using a load balancer, and optimizing inference with batching, quantization, or a faster runtime like ONNX. Mention caching for repeated queries.

3. 7

   A data scientist gives you a notebook that trains a model. How do you convert it into a reproducible pipeline?

   Extract code into modular scripts, parameterize inputs, use a workflow tool like Airflow or Kubeflow, version data and code with DVC or Git, and log metrics and artifacts with MLflow. Ensure the pipeline can run end-to-end with one command.

4. 8

   How would you monitor a deployed model for performance and fairness?

   Track latency, throughput, error rates, and prediction distributions. For fairness, monitor metrics across subgroups and set up alerts for significant disparities. Use dashboards with Prometheus and Grafana, and log predictions for periodic audits.

Behavioural and HR

1. 9

   Tell me about a time you had to learn a new tool or technology quickly for a project.

   Use the STAR method. Focus on how you identified the need, your learning approach (docs, tutorials, small experiments), and the outcome. Show curiosity and self-drive, which are valued in fast-growing AI companies.

2. 10

   Describe a situation where you worked with a team to solve a technical problem. What was your role?

   Highlight collaboration, communication, and your specific contribution. Mention how you handled disagreements or blockers, and the final result. Emphasize that you are comfortable working with data scientists and engineers.

3. 11

   Why do you want to work as an MLOps Engineer at a smaller AI company like Wingman?

   Show that you understand the company's focus and the broad ownership MLOps engineers have in smaller teams. Connect your interest in building end-to-end ML systems and wearing multiple hats to their environment.

4. 12

   How do you prioritize tasks when you have multiple urgent requests from different teams?

   Explain assessing impact and urgency, communicating with stakeholders, and using a simple framework like Eisenhower matrix. Mention keeping a visible task board and being transparent about trade-offs.

Coding and Tools

1. 13

   Write a Python function to read a CSV file, handle missing values, and return a cleaned DataFrame.

   Use pandas, show awareness of dropna or fillna, and mention handling data types. Keep it simple and readable, and explain your choices for missing value treatment.

2. 14

   How would you use Git in an MLOps workflow? Mention branching and versioning of data and models.

   Describe feature branches, pull requests, and tagging releases. For data and models, mention DVC or Git LFS to avoid bloating the repo, and linking model versions to Git commits for traceability.

3. 15

   Given a SQL table of model predictions with timestamps, write a query to find the average prediction score per day for the last 7 days.

   Use DATE_TRUNC or CAST to day, filter with WHERE timestamp >= CURRENT_DATE - INTERVAL '7 days', and GROUP BY day with AVG. Show correct syntax and mention timezone considerations.

4. 16

   What is the difference between a Docker image and a container, and why does it matter for ML deployment?

   Explain that an image is a immutable template and a container is a running instance. For ML, images ensure consistent environments across training and serving, reducing dependency issues. Mention layers and caching for faster builds.