The ElevenLabs onsite is 4 rounds:
- V1: Live coding
- V2: Product decomposition
- HM
- CTO
The V1 problem was wrapped in an AI shell. The text generated by the LLM is want_to_speak, and what the agent actually says is spoken_text. want_to_speak is always correct, while spoken_text may contain repeated characters, words or word phrases. You have to return the longest prefix of want_to_speak.
For example:
want_to_speak: I can help you book an air ticket
spoken_text: I can can help you book
expected output: I can help you book
The first part is reviewing a junior dev's code. He used LCS: first build a DP matrix to get the LCS length, then rebuild the LCS from right to left. There were two mistakes. One was that his DP matrix index goes out of range. The other was that this case doesn't work:
want_to_speak: I can help you book an air ticket
spoken_text: I can can
output: I can an
expected output: I can
The second part asks you to implement function_v2. Reading the problem carefully matters a lot, and I only figured it out near the end: spoken_text only ever has repetitions, it never drops words or substitutes words. So it is really just two pointers moving left to right, and whenever the two don't match you only advance the spoken_text pointer.
V2 was really open-ended. For example, take the Foot Locker website and design a voice chat system, which is extremely open. I kept asking the interviewer what they thought it should look like, and they would just reply with what do you think it should look like. I don't know whether that was one of the things that sank me. The feedback was: we felt there were gaps in AI fluency and product design skills that are critical for this role.
Discussion
Loading comments…