Cohere · Primly Community

Cohere software engineer interview process, full loop: here's what I went through in 2026

market_realist · 4 replies

Just finished the Cohere software engineer interview process and wanted to write up the full loop while it's fresh. Applied through LinkedIn, heard back in about 10 days, which felt unusually fast compared to my other apps.

The sequence was: recruiter screen (30 min) -> technical phone screen with an engineer (45 min) -> take-home or async coding challenge -> onsite (4 rounds, all virtual). Total calendar time from first contact to verbal offer was about 5 weeks.

Recruiter screen was pretty standard: background walk-through, why Cohere, compensation range check, visa status. She was friendly and gave a good overview of what the loop looked like, which I appreciated.

Technical phone screen had two parts. First 20 minutes was algorithms, a medium-difficulty problem about string parsing. Second part was a conversation about a past project, pretty deep technical questions on design tradeoffs I'd made. Not a leetcode grind session, more like they wanted to see how you reason through things.

Onsite (4 x 60 min) covered: Coding (two problems, medium-ish, Python or Go preferred but they said any language was fine) System design, specifically LLM/inference infrastructure adjacent. More on this below. Behavioral, 4-5 questions, standard STAR stuff but they pushed hard on the "what would you do differently" layer Cross-functional interview with someone from product or research, more conversational

The system design round was notably different from what I'd seen at other AI companies. They were genuinely interested in how you'd think about serving large language models at scale, token budgets, batching, caching inference results. You don't need to be an ML expert but knowing vaguely how transformer inference works helps a lot.

Overall Cohere felt more collaborative than evaluative. Interviewers seemed like they were having an actual conversation rather than running through a script. Zero puzzles, no trick questions.

I got the offer. Happy to answer questions.

4 replies

jp_newgrad

thank you for writing this up!! quick question, was the take-home graded or did it just unlock the onsite? like did anyone ever mention it during the onsite rounds?

market_realist

good question. nobody referenced it directly during the onsite, but I think it was a pass/fail gate to get to the onsite rather than something they scored deeply. my recruiter said they use it to check code quality and structure rather than grinding on edge cases.

ml_mike

The inference/serving angle in the system design is interesting. Did they want you to go deep on like KV cache, speculative decoding, any of that? Or more high-level architecture of a serving cluster?

market_realist

More high-level. We talked about horizontal scaling of inference workers, request routing, batching strategies. KV cache came up but I didn't need to go into the math. It was more about whether you've thought about this class of problem before.