Updated by Meta candidates
Computer Vision Engineer (L5) Interview Experience
What caught me off guard was that one of the coding rounds was not generic LeetCode at all. It was AI coding, with things like sparse matrix times vector multiplication and k-nearest neighbor search, and they still wanted multiple questions under pressure.
Interview process
What surprised me most was how much the process emphasized solving several coding questions under pressure. I had an initial recruiter screen, then a 45-minute pre-screen with a senior manager about my past work and impact, and after that a full loop of five interviews. The loop was two coding rounds, one of them classic LeetCode-style and one of them AI coding with data-processing and math-heavy prompts like sparse matrix times vector multiplication and k-nearest neighbor search, plus two design rounds and one behavioral. The design interviews were split between an ML modeling discussion for a computer vision problem and a system design round on something like Google Street View, which I found pretty challenging. I passed the loop, got notified that I made it through, and then moved into offer negotiation.
- Recruiter screen
- Phone interview
- Technical interview
- Other
Interview rounds
10 questions across 7 roundsRecruiter screen
1 questionOverview
I started with a recruiter screen that was basically a fit check for whether my background matched the role.
Question types asked
Specific questions asked
Phone interview
1 questionOverview
After that I had a 45-minute pre-screen with a senior manager, and it was very focused on my past work, what I had actually delivered, and what impact it had.
Question types asked
Specific questions asked
What exactly did you do on each project?
What was the impact of the things you did?
I went through my past experience project by project and focused on what I had delivered and the impact. The tone of that round was very much about whether I had actually done the work and could speak concretely about it, not just give a high-level summary.
Technical interview (1)
Overview
One of the coding rounds felt like a more classic LeetCode-style interview. The pressure point was the pacing, because they told me upfront we might do two or three questions depending on time.
Question types asked
Technical interview (2)
3 questionsOverview
The other coding round was the most distinctive part for me because it was not generic LeetCode. It was what they called AI coding, so the questions were more about data processing and math that felt typical of AI work, and they were still pushing for two or three problems under time pressure.
Question types asked
Specific questions asked
I treated it like an AI-flavored coding problem rather than a standard data structures one. The notable part was that I had to code it efficiently under pressure, knowing they wanted multiple questions in the same round.
Again, I approached it as a data-processing and math-heavy coding prompt. What stood out was less the exact algorithm and more that this round clearly tested whether I could handle AI-style coding fast, not just do generic interview problems.
What's the general idea for the third question?
In one of the coding interviews, they were pushing on pace enough that for the third question the interviewer just asked me for the general idea instead of having me fully code it. That gives you a sense of how compressed and speed-focused these rounds were.
Technical interview (3)
1 questionOverview
One design round was really a computer vision and ML modeling discussion, not a traditional system design. High-level: discussion centered on depth from a single image and involved real back-and-forth on my approach.
Question types asked
Specific questions asked
How would you approach depth from a single image?
Can you go deeper on the nuances and trade-offs in your modeling idea?
I walked through my idea and then went back and forth with the interviewer on it. It was a modeling discussion with nuance and pushback rather than a generic open-ended design chat.
Technical interview (4)
1 questionOverview
The second design round was a true system design interview and I found it challenging to crack. The prompt was to design something like Google Street View, and the hard part was balancing the algorithm side, the collection side, and how the whole thing would be handled in production.
Question types asked
Specific questions asked
How would you think about the collection side?
How would you handle it in production?
I had to think about it as more than just the algorithm. I tried to balance the algorithmic piece with data collection and production handling, and that was what made the round challenging because they were clearly looking for a full end-to-end design, not just one strong technical slice.
Other
3 questionsOverview
There was also one standard behavioral round. It was the usual big-tech style set of stories around pressure, conflict, and growth after tough feedback.
Question types asked
Specific questions asked
Interview tips
Generic LeetCode will NOT cut it. I would practice doing two or three coding problems back-to-back under time pressure, and I would make sure some of that prep is AI-flavored coding around data processing and math, not just classic algorithms. I would also be ready for two very different kinds of design interviews: one that is really ML modeling for computer vision, and one that is true end-to-end system design. For behavioral, I would have clean stories ready on delivery pressure, conflict, and responding to tough feedback.
Company culture
The process felt pretty standard for a big tech company, but with a heavier emphasis on fast live coding than I expected. The interviewers were also pretty deliberate about pacing and scope, so it did not feel like casual conversation rounds. It felt structured, high-bar, and very intentional about covering both modeling depth and production thinking.
Get full access with a membership, or share your experience to try it free.
More Meta interview experiences
“The AI enabled round was actually pretty unique. I used bit mask first, then hit a collision because the input had letters and numbers, and the LLM started hallucinating new functions, so I had to debug both the codebase and the AI.”
“I got feedback that I performed well on the coding interviews even when I did not feel like I performed up to a requisite level, so I think the bar has more flexibility than we may give ourselves credit for.”
See all 13 Meta Machine Learning Engineer interview experiences →
