Cohere ML / system design interview

Round 5 of 7 · 8 questions reported

See the whole loop
Length
60 min
Format
Design discussion
They grade
Enterprise concerns: multi-tenant RAG, isolating customer data, VPC and on-premise deployment, cost per query, latency, GPU scheduling

Questions Cohere asks in this round 8

  • Design a real-time comment feed with optimistic updatesSystem designDesign the front end of a comment section where other people's comments arrive live and the viewer can post their own. The focus is optimistic posting: a submitted comment shows…
    Last asked Sep 2026
  • Load daily LLM chat logs into a warehouse idempotentlySystem designEvery day one JSONL file of LLM conversations lands in object storage; each line is a conversation with an id, a user id, its messages and metadata. Design the job that loads it…
    Last asked Sep 2026
  • Defend your own research paper under hard questionsML designYou present one of your published papers to an interviewer who has already read it and may skip your slides to go straight at its weak spots. Be ready to say whether the…
    Last asked Sep 2026
  • Make an LLM recognize questions past its training cutoffML designChat assistants sometimes answer a question about recent events by saying their knowledge stops at a certain date. You are building an LLM: how would you get it to notice that a…
    Last asked Sep 2026
  • Post-train a 7B model to code better for one enterpriseML designAn enterprise customer wants a 7-billion-parameter model that writes better code for them. Design the post-training pipeline end to end: collecting and cleaning data, supervised…
    Last asked Sep 2026
  • Improve an LLM's reasoning on a domain taskML designA model reasons poorly on one specialized task. How would you improve it? Compare the options, from prompting and retrieval to fine-tuning and reinforcement learning, say what…
    Last asked Dec 2025
  • Design a RAG system over company docsAI systemEmployees should be able to ask questions about internal documents (Drive, Confluence, Slack, GitHub) and get grounded answers with citations.