1 phone screen, 1 x 1-hour pair coding round (AI-assisted coding round), 4 x 45 min rounds in one day.

The first round was with HR (Toby). It was generally to introduce yourself and the role. The role focused on the User Data Platform group, and the core requirements were expertise in the NLP and NLG space, along with knowledge of fine-tuning and prompt engineering using popular open-source LLMs.

The second round was the AI-assisted coding round. It was so vague – I had absolutely no idea what I was supposed to do. I asked the HR as many questions as possible. He was very vague about everything: Was this going to be a DSA round? A system design round? An ML problem-solving round? No idea. The only thing that was made clear was that I would be presented with a business problem and was expected to create a solution using my development environment (unit testing frameworks MUST be installed previously; AI-assisted coding tools like Cursor or GitHub Copilot must be installed previously), and I was expected to write test cases. The kind of code I write should be production-level and clean. My "prompting" would be highly assessed. I was expected to create the project quickly with AI tools.

What would you expect? Some ML problems and the expectation to create a clean project, good project structure, and test cases, right? WRONG.

The interviewer came and asked me a business problem around "search" in Canva – specifically, gender bias in Canva's search. Why? Because he belonged to the team handling that! How does that even make sense – evaluating senior candidates in domains they have no experience in? Just because the interviewer happens to work in that area? And it's not even relevant to the JD.

Anyway, I tried to tackle it as a general machine learning problem. I suggested techniques on improving the training dataset, talked about dataset balancing, data augmentation, even prompt engineering approaches for retrieval enhancement, and other training models.

But the interviewer asked me to shift to NON-ML-based approaches. This is a search infra/ranking fairness problem involving image retrieval – not something remotely close to NLP/NLG, prompt engineering, or fine-tuning LLMs.

And during no part of the interview did the interviewer want to see my production-level coding skills, unit test cases, clean code, use of AI-assisted tools, or my prompts. It felt like he would’ve been happy if I just wrote it in a ChatGPT window – maybe only the algorithm.

Of course, I was able to come up with a solution with hints, but it’s naturally going to be biased against someone who doesn’t come from this domain – especially when it had nothing to do with the JD.

The thing is, I was so surprised and didn’t expect this at all that it took my brain at least 15 minutes to understand that I didn’t have to build a coding project and just solve it as a general business problem. I honestly think I would have been better prepared if I hadn’t received any information about the coding round at all. It was so confusing.

Someone coming from a similar domain would probably know more than me, but it’s just not aligned with the JD. That’s unfair – just because the interviewer on that day belonged to that specific team.

You are an engineer working on search. There has been an incident on social media: users are upset that photo results for certain queries have poor gender diversity. For example, “bodybuilder” shows mostly male bodybuilders. More and more users are posting new queries where results lack gender diversity.

This is occurring because our content library is skewed: for each female bodybuilder image we have, we might have 50 male bodybuilders.

Another team is manually elevating certain results for certain queries, but this is a manual process and will not scale well.

You have been asked to ship a more general mitigation for this problem by end of day. What would you do?

MMMMM, FFFFFF. He asked me to write an algorithm for indexing, outputting the results in a non-biased order (Round Robin). Also, to suggest evaluation metrics for this (e.g., Diversity Index like the Shannon Index).