Your goal: design a chat service you can size, speed up and ship in stages, and make a call
when saving money costs quality.
How this works (about 30 minutes)
- Felix asks which role you are interviewing for (product, engineering or other) and pitches his
follow-ups to it. The question stays the same.
- One design question, taken deep. Ask him about users, traffic, models and goals. He hands over
what he has, one line at a time, and later a latency and cost sheet. All numbers are made up and
illustrative; none is a real price or speed.
- Many follow-ups, one curveball, then a trade-off turn from your own routing choice.
- If you want to sketch, say "can I use the notepad?" and Felix puts it up.
- No hints and no scores during the call. Your report comes after.
What scores well
- Questions about users, traffic, latency targets and quality before you draw
- A request path with streaming, state, and capacity sized in tokens, not only requests
- A model choice by request type, with a router and an escape for its mistakes
- Caching and batching with their costs named
- A staged rollout with slice gates, a stop rule and a way back
- A call you can defend, and what would change it
- Answers that say where they are going and land in a minute or two
Tips
- Take Gen AI Coach · Shipping LLMs first.
- Say your numbers aloud and label them as assumptions.