Anthropic Software Engineer Interview Experience — Onsite Loop with a 1-on-1 Chat System Design, File Dedup Coding, and a Culture Round I Couldn't Read

The phone screen was Q1 image processing — using PIL to do grayscale, scale, and resize. The first step handled small images, the second step handled large images, and improving performance meant using ProcessPoolExecutor. The recruiter had mentioned beforehand that I should be comfortable with parallelism/concurrency, which was basically a hint. Time was a bit tight and the interviewer had to nudge me before I finished, but overall it went okay and I passed.

Onsite had a few rounds, so I'm noting them down here.

The system design round was interesting — I had to design a 1-on-1 chat system, and the interviewer stressed that it was only 1-on-1, group chat and channels were out of scope. I hadn't seen this exact question on the forum before, so I'm not sure if it's new or if the interviewer just felt like asking it that way. When I was prepping I'd looked at a chat system design writeup — the original problem was Slack's design with groups and channels, and you just strip that part out for 1-on-1, the core ideas carry over. We talked about how to push messages when the recipient is offline, how to know when a user comes online, how sessions get stored, and then it moved into Kafka vs Redis and how Kafka actually works under the hood — each part got probed pretty deeply. Looking back, I never fully understood Kafka internals, and I was on the back foot once he started digging.

The coding round was file dedup — deduping by file size and content hash. The problem itself wasn't hard, but the follow-ups went deep: how do you tell if something is IO-bound vs CPU-bound, how do you handle huge files, what about a scenario with a massive number of files, what mechanism would you need for realtime detection. I fumbled a few of the follow-ups — the base problem was easy but I ended up dying on the follow-ups instead.

The culture round was harder to read than the technical ones. The interviewer asked why Anthropic, and wanted a specific value plus a real example from my own experience, not just a vague "I'm interested in AI." He also asked about my views on AI safety and any moral dilemma I'd run into. The whole time I couldn't tell what score he was giving me or what kind of answer he actually wanted.

The HM round was more relaxed, mostly about a complex project I'd worked on and its roadmap impact, and how I mentor other people. The interviewer was a bit cold though — forty-plus minutes in and it felt like I was doing a solo stand-up routine the whole time, hard to tell which direction he wanted me to go.

Now I'm waiting on the result. My guess is the culture round is my weak spot — hoping it goes through anyway.