How ScaleCrafts grades system design
System design has many valid answers, so ScaleCrafts doesn’t check your diagram against a single “correct” one. It checks whether your design handles the hard parts of the problem, the way a good interviewer would.
1. Design it
Every problem lists functional and non-functional requirements and the scale to plan for. You draw the architecture on a canvas (clients, load balancers, services, caches, databases, queues, workers and more) and connect them. Calls can be synchronous or asynchronous. Components carry the settings interviewers ask about: how many instances, the shard key, and a queue’s delivery guarantee.
Notes follow the order of a real interview: requirements, estimates, API, data model, and design and deep dives.
2. Get scored on concerns, not shapes
Each problem’s rubric lists its hard concerns. For a ticketing system, that’s never selling the same seat twice and surviving an on-sale rush. Each concern accepts several named approaches, such as a database transaction, a lock backed by a unique constraint, or processing each event’s purchases one at a time. When you pass, you see which approach was recognized.
- Mistakes cost points. A client talking straight to the database, a single instance on the critical path, or components that aren’t connected all lose points.
- Drawing everything doesn’t help. Components you never explain count against you. Explain each one in your notes, or in its own “What does it do?” line on the diagram.
- Common mistakes show up as you draw. The checklist under the diagram flags these before you submit. Your problem’s own rubric stays hidden until you’re graded.
- Numbers matter. Estimates need real figures with units, like “40k QPS” or “5 TB”, not just the word “QPS”.
You need 70% to pass. Hints appear for anything you miss, and you can resubmit as often as you like.
3. Ask the AI reviewer
After grading, you can ask for a written review of your submission and then ask follow-up questions about it. The reviewer only discusses your design. Every point it makes must cite a component, connection or quote from your notes, and those citations are checked automatically. It can’t change your score.
4. Take a mock interview
For the full experience, start a timed mock interview on any problem (45 minutes by default). An AI interviewer asks one question at a time about requirements, estimates, your API and data model, scaling, failures and trade-offs, while you keep drawing. It sees your diagram and notes as they are when you answer, and it asks about your design, not a generic one. It never tells you what to add.
- Graded at the end. When time is up, or you finish, your design is graded exactly like a normal submission.
- A debrief. You get scores from 1 to 5 for requirements, design, depth and communication, what went well, what to work on, and how each answer was rated.
- Same guard rails as the AI review. It stays on the interview, and every point it makes about your design is checked against your diagram and notes.
Start practicing
Browse all problems, or start with an easy problem such as a URL shortener.