Scale AI New Grad Problem Bank: 30 Questions and Practical Projects (2019–2026)

Coverage: New Grad SWE and MLE · OA, phone screen, onsite coding, debugging, backend practical, ML practical, system design, and low-level design · 2019–2026

Speical thank you to a fren for their patience~

The short version

The question bank is broad, but the recent SWE process puts a lot of weight on practical implementation.

1. The current phone screen is usually one long build. The task grows in parts, so clean code matters as much as getting the first part right. The task processor and four-player card game are the two strongest recent examples.
2. The current onsite has two technical rounds. Day 1 is a one-hour debugging round and a one-hour backend practical. The contractor assignment repo has appeared more than once, and the same setup has stayed very similar across years.
3. The OA can still be algorithm-heavy. Arrays, dynamic programming, intervals, trees, graphs, and simulation all show up.
4. MLE has a different problem mix. Expect data loading, model evaluation, debugging training data, tree or forest distance work, statistics, and small ML pipelines.
5. Older questions are still useful, but newer ones should come first. The best order is the 2026 phone-screen and practical work, then the repeated graph and simulation questions, then the older OA set.

A practical prep order: task processor → card-game simulation → contractor debugging repo → backend service → repeated graph and forest questions → older OA problems.

Start here for the 2026 SWE loop

1. Task Processor: Dependencies and Deadlines — Seen: Aug 14, 2026; Jun 14, 2026; Apr 30, 2026; Mar 27, 2026; Mar 12, 2026
2. Four-Player Trick-Taking Card Game Simulation — Seen: Aug 31, 2026; Dec 1, 2023
3. Repair Contractor Project Assignment — Seen: Apr 21, 2026; Oct 31, 2024
4. Build a CSV-to-JSON LLM Categorization Service — Seen: Aug 1, 2026; May 9, 2025
5. Format and Validate a Model Response — Seen: Apr 30, 2026

The first two match the long, multi-part phone-screen style. The next two cover the current debugging and backend practical format. The last one is a good short exercise for parsing, validation, and careful output handling.

Online assessment: 9 problems

Two OA prompts are especially easy to miss:

• Count Stable Server Subsegments: count contiguous segments with at least three values where both endpoints equal the sum of everything between them.
• Distinct OR Values of Increasing Subsequences: return every different bitwise-OR value that can be made from a strictly increasing subsequence. Include 0 for the empty subsequence and sort the result. The reported value bound was below 1024.

Phone screen: 8 problems

The important pattern is not one data structure. It is handling a task that keeps growing. A clean model, small helper functions, and quick tests make the follow-ups much easier.

Seen in both phone and onsite rounds: 2 problems

The forest problem is one of the most repeated items in the whole bank. The poker problem tests careful rule handling and edge cases more than a hidden trick.

Onsite coding and ML: 5 problems

These range from basic array work to stateful matrix updates, graph distance, response validation, and numerical estimation. For MLE roles, being able to explain the math and test the implementation matters just as much as finishing the code.

Debugging and backend practicals: 4 projects

The contractor project is the main debugging practice for the current loop. It uses an existing codebase with contractors, courses, projects, and CSV data. The bugs sit in real program behavior: capacity checks, priority order, course filtering, and state that is not reset correctly.

The backend project is different. It asks you to build working behavior across several parts, not just find planted bugs. Practice reading the existing structure first, then make the smallest change that gets one requirement working and tested.

System design and low-level design: 2 problems

These are more role-dependent than the coding and debugging rounds. The first covers an end-to-end retrieval and generation system for claims work. The second focuses on objects, assignment rules, state changes, and a clean model for tasks and workers.

This bank will be updated when new firsthand reports add a new problem or a useful follow-up.