After the initial HR call I got an invitation for a 3.5 hour long on-site technical "interview" in the middle of a working day (no late evenings or early mornings hours were available for taking the test). I got to the interview and the person that was suppose to interview me was sick, and none of the other office workers had a clue I'm coming. After about 20 minutes of waiting somebody eventually set up a computer for me to take the test on. The test had 3 parts: 1) leetcode style programming task (easy) 2) 10 open questions about LLM, RAG, and vision tasks. 3) two "code filling" tasks: task 1: An LLM is prodoucing wrong/inaccurate short answers. the task is to examine the prompt for a LLM and to to tweak it hyperparameters (temp, num of tokens etc') task 2: you are given a "legacy" code snippet that trains a pytorch vision classification mode, the code is messy, incorrect and lacks any documentation. Moreover, the trained model performs badly. the task is to rewrite / fix the code, examples of errors in the code: - the training val loader is not used for validation (they used the train loader) - the mean loss calculation for each epoch was wrong. - the model was not set to "model.eval" before validation. - no metric logging was used. After all of that you are required to fill in a survey about the "interview process" where they ask you about the "maximum appropriate length of the interview in your opinion" and such other stuff, and ask if you would agree to take ANOTHER such interview in the future. I finished all of the tasks, nobody talked to me, I left the company offices alone and after about two weeks I got a standard rejection call.

The interviewing process was smooth and clear from the start, it consisted in phases, first HR, then a technical interview with the Team Manager, then a technical interview plus take home coding exercice with the final client.