Figma Machine Learning Engineer Interview Experience — Agent Post-training Phone Screens and a Colab PyTorch Pipeline I Could Not Finish

Sharing my Figma MLE interview experience (Agent Post-training).

A recruiter reached out to me in mid-September. The role is mainly about Agent-related post-training work, and I then scheduled a half-hour recruiter call.

One: Recruiter screen (30 mins)

The recruiter described the background of the role and what the work involves, then asked about my background and a few common BQs, including:

• Walk through an end-to-end project you owned
• Why are you considering leaving your current company
• Your long-term career goal, and what direction you want to pursue going forward

At the end they also asked about my salary expectations and my immigration status.

The day after the recruiter call, I got the schedule for two phone screen rounds.

Two: First phone screen, HM round (30 mins)

This one was mostly about my background, the scope of my work, and the hands-on projects I was responsible for. The interviewer dug deep into project details, for example:

• How did you make the specific technical decisions in the project, and why did you pick that approach?
• What strategy did you use?
• How did you design and run the evaluation?
• Other specific technical implementation details of the project

Overall this round was mostly chatting and project discussion. There were no ML trivia questions, it was mainly a deep dive into projects I had actually done.

Three: Second phone screen, ML coding (1 hour)

There was about 5 minutes of self-introduction at the start, and then we went straight into the problem. The problem was in Google Colab, where I could edit the code directly and run tests. It used PyTorch, I was allowed to look up PyTorch usage, and I had to share my screen the whole time.

Problem background: there is a large amount of text description that needs to be annotated, marking the highlight snippets in the text. Training, validation, and eval sets were provided. Each sample corresponds to a text section and comes with:

• the original text (the text section)
• the text snippet that should be highlighted (the highlighted text)
• the start and end position index of that snippet in the original text

Requirement: based on the existing code skeleton, fill in a complete training pipeline, including model definition, setup, forward, training, and evaluation. Each part had a skeleton provided, but the actual implementation details were for me to fill in.

The problem as a whole was fairly long, and the problem statement was worded in a convoluted way, so just understanding the task and sorting out the code skeleton took me a lot of time. In the end I only finished the model definition at the beginning, plus the forward-related parts of training/evaluation. There was still quite a lot after that which I did not get to look at, let alone finish.

Summary

My experience so far is that the HM round tests real project experience and technical decision-making, while the ML coding round is a fairly complete PyTorch training pipeline implementation, not just writing a single algorithm or implementing one simple model.