ML system design interview practice
Design one production ML system on a shared whiteboard, such as a recommender or a fraud detector. Vera asks what the model predicts, where the labels come from and what breaks after launch.
- 45 to 50 min
- One ML system, end to end
- Up to two small nudges
- Shared whiteboard Vera can read
How the round runs
- 01
Requirements and scope
Vera states the problem in a sentence or two and invites questions. Pin down the goal, what the model should predict, the scale and the latency budget before any boxes.
- 02
High-level design
Walk through the flow end to end: data in, training, the model, predictions out. Expect questions about the data model and storage too, because this round still uses the general design stages.
- 03
Deep dive
Vera picks the riskiest or least explained part, often the features, the labels or the model choice, and asks how it actually works here.
- 04
Scale and failures
One twist at a time from the question, for example a new item with no history, offline metrics that improve while the A/B test stays flat, or latency that doubles after new features.
- 05
Wrap-up
Name the weakest parts of your design, how you would monitor it and roll it out, and what version two would add.
What the report scores
Each skill gets a score from 1 to 5 against the bar for your level, with the quotes it is based on and the time you said them. The report ends with one thing to fix next.
- Requirements and scoping
- Structure and driving
- API and data model
- Architecture
- Depth
- Trade-offs
- Scale and failure handling
- Communication
At lead level, also: Operability and cost, Scope and influence.
How the real round works
An ML system design interview usually lasts about 45 minutes and follows the life of a model: turn the business goal into an ML objective and a label, sketch the pipeline, work through data and labels, features and a model, evaluation offline and online, then serving and monitoring. The guides we read agree that it grades ML judgment rather than infrastructure, and that opening with load balancers and REST endpoints is a named mistake.
How ML system design interviews work: structure, signals and common mistakesRegular and lead level
- Regular
- Vera guides gently. Jump straight to a model and you are asked for the requirements and scale first; stall and you get a small prompt about the next area.
- Lead
- Vera does not structure it for you. You drive the agenda, pick the risky parts and cover cost, operations and team impact. Name a technique and you will be asked how it works here.
Length: 45 minutes at regular, 50 at lead.
See the feedback before you start
Full reports from the real report pipeline, for a fictional student and a scripted interview.
Questions
Is it scored differently from a backend system design?
Not yet. Today the ML round uses the same five stages and the same skills as every system design round, so the report also scores the API and data model. The questions, their clarifications, deep dives and twists are ML ones.
What kind of systems come up?
Recommendations, search and ads ranking, fraud and abuse detection, content moderation, forecasting, and LLM systems such as an assistant over a company’s documents or a model serving platform.
Do I need to know a specific framework or cloud?
No. The round is about the decisions: the label, a baseline, how you evaluate offline and online, and how you serve and monitor the model. Name a tool only if you can explain why it fits.
Is the sample report an ML design?
The linked sample is a general system design round. It uses the same stages and scorecard the ML round uses today, so it shows what your report will look like.