Scale AI New Grad SWE Interview Experience

Role: New Grad Software Engineer
Location: United States
Result: Rejected
Difficulty: Difficult
Process length: About two months

Cold applied. First step was a 50-minute HackerRank, with no recruiter screen beforehand.

The HackerRank used an interval problem inside a more practical API-style setup. Party schedule data had to be read through a provided method instead of a plain array. The task was to find the dead zones between party time intervals. Finished in about 20 to 25 minutes.

After a gap of a few weeks, the next step was a 60-minute technical screen with two interval problems.

The first problem asked for free time ranges where new queries could be accepted. A follow-up asked how to order the intervals to reduce the amount of time spent processing queries with an LLM.

The second problem was similar to Meeting Rooms. Given schedule data, return which rooms were free and the time ranges when they were available. Input again came through a provided method rather than a direct array.

The main difficulty was speed. The interval logic was manageable, but both solutions had to be written cleanly within one hour. The round was candidate-driven, with help available mainly for clarification.

Next came a system design round and a debugging round back to back.

The system design question was an insurance-claims agent. Emails and PDFs containing claim information would be ingested, relevant data would be extracted with RAG, and the system would return an approval or denial decision.

Discussion included storage for policy documents and customer policy data, infrastructure for document ingestion, and ways to control LLM token cost.

The debugging round provided roughly 150 to 200 lines of modular code and an API key for a Scale-hosted model endpoint. The program was not returning model output.

Print statements were used to trace the active methods and narrow down the failure. The bug was found in the prompt-hashing flow. Hashing was being used to reduce token cost, so the fix required understanding how that path worked rather than removing it.

Rejected after the system design and debugging stage. No detailed interview feedback was provided.