The first stage involved a take-home machine learning assignment on a 12-way text classification problem. A potential solution is to use character-level CNN for text classification.

SAP only invites candidates who score more than 70% on the test set. I learned this by asking one of the interviewers in the second stage.

The second stage involved two consecutive rounds of in-person interviews at the Singapore SAP office. Each round lasted one hour and had two interviewers.

In the first round, each interviewer asked one whiteboard programming question. The questions were on merging two sorted lists and how to query a large database that keeps growing.

The interviewers also asked more questions on NLP, as that was the area they were working on at SAP. They did not seem to care about the type of machine learning projects you have worked on, simply asking questions pertaining to their own projects at SAP.

In the second round, I was interviewed by two different interviewers. One interviewer asked a whiteboard question on masking in the TensorFlow API, while the other did not ask any whiteboard questions.

One of the interviewers clearly informed me that they were looking for a wholesome candidate and wished to avoid "silo candidates" who are only experts in one area, such as machine learning. He wanted someone knowledgeable in the full software development cycle across the entire full stack.

In summary, the interviewers appeared to be unprepared for the interviews. They had not read my resume before the interview, and I had to re-introduce myself to each of the two panels.

Most questions were asked on the fly without preparation. Truthfully, the interviewers were actually looking for someone who is a good full-stack software engineer, as evidenced by the question on querying a database and the "non-silo candidate" comment.

Although this job was advertised as a machine learning role, you are highly likely to get this job if you are a good full-stack software engineer with very little actual machine learning knowledge.

Merge two sorted lists into one sorted list.

How do you perform masking in the TensorFlow API?

How do you query a large database efficiently, knowing that the database keeps growing over time?