Scale AI Backend Engineer Interview Experience — Full Onsite Loop with a Claude-Assisted Coding Round

I mass-applied for an infra role that sits under the ML research org. The interview process matched what other people have already posted here on the forum.

Onsite loop:

• Task scheduler

• VO

• BQ: Standard questions, and it ended early. The highlight was during the time to ask questions back — I asked the interviewer if there was anything she didn't like about Scale AI, and she said that because of RTO3, a lot of people don't come into the office, so it's kind of lonely there.

• Technical deep dive: Give a presentation introducing one of your own projects. One thing to watch out for in this round: the interviewer is a general SWE, so if your work is very domain-specific, you need to prepare your slides to explain it in easy-to-understand concepts, and it needs to be metrics-driven.

• System design: The embedding + classification API question (probably every AI-related role gets asked this one). What's kind of wild is that when I described this question to Claude, it could go find the original problem directly and even knew it was a Scale interview question.

• Backend practical: Two questions. The first is to write an HTTP server that takes a CSV path and writes a JSON file to the filesystem. The second is to add a category field to that JSON, where the category is generated by calling an LLM API (they give you the key for this). The follow-up is how you'd scale it up.

  I really want to vent about this round. It's billed as "AI enabled," and I thought it would be interesting, but from the format it's really just the pre-AI-era version of the question, and they just let you use AI on top of it. After I double- and triple-checked whether I could use "any" AI, I originally wanted to have Claude's web version OCR a screenshot of the question into text and then just paste the prompt straight into Claude Code, but I was stopped — they said I had to at least write the prompt myself... lol.

  Then, naturally, powerful Claude solved it in seconds — I didn't even need to turn on Max. For the follow-up, I discussed sharding and calling the LLM API in parallel with the interviewer. The interviewer said, "okay, write me a parallel call then," and when we looked at the code, Claude had already written exactly that version (laughs). The whole round wrapped up in under 20 minutes.

• HM: Mostly just chatting — telling stories, listening to the hiring manager talk about what projects are on the team. No technical questions at all.

Overall the difficulty was medium-low.