I was approached by a Mistral employee on LinkedIn. I was not actively looking for a job at the time, but felt it was worth trying.

First, I had a call with this person.

I was then sent various links to resources to best prepare, mostly about LLM evaluation.

Second, I had a Python interview focused on refactoring. The interview was not hard for someone with experience in Python, but at times the interviewer did not seem very attentive. For example, when I asked if I should use sync or async logic, they asked me in return if I knew what async was. Several times, they also asked questions and answered them right away without really waiting for my answer.

Third, I had a chat about LLMs with another Mistral employee, mostly about RAG. The interviewer was nice and open-minded, though at times I could not get them to justify their choices. For example, when discussing RAG versus fine-tuning for a specific use-case, the interviewer favored fine-tuning because "it works better with this use-case," with little further elaboration.

I heard nothing from them for about 10 days, then was invited to the fourth step. Their interview scheduling system is a bit unreliable; most of the time, there are no availabilities, and you have to regularly check the portal to hope to book a time.

Fourth, I had a Python pair programming interview. The interviewer was late and had network issues on their side (not their fault, but worth noting for context). The exercise was about using a third-party API for information retrieval in combination with the Mistral API, in order to add the external information/data to the context. The exercise was easy enough to understand. The interviewer made me share my screen to see my IDE. Their network issue made it difficult for them to follow. I could feel they were losing patience at times, and they even had to disconnect from the call a couple of times because of it. In the end, I managed to complete the exercise, even though we could not test the end-to-end result because we did not have a Mistral API key available, making the end goal of the interview a bit unclear. The interviewer wrapped up the interview due to lack of time and told me they would be in touch.

I received an email the next day stating the interview process would not continue because my skills with APIs were not meeting their expectations.

Anyway, without being rude to anyone, most of this process felt like a waste of time. Interviewers were nice people overall, but speaking as someone with experience in tech recruitment, it did feel like each individual step's assessment goal was unclear at best.

Describe what a RAG is.

If you have a 2000-page PDF document (i.e., something that does not fit in the context window) that serves as a knowledge base, would you rather fine-tune a model or set up a RAG?