Ebay ML Engineer interview typically runs 4 rounds: technical director, model training and inference, coding assessment, final technical. It usually takes about 4 rounds and leans heavily into low-level GPU and kernel work.
$163K
Avg. Base Comp
$228K
Avg. Total Comp
4
Typical Rounds
2-4 weeks
Process Length
We’ve seen eBay’s ML Engineer interviews reward candidates who can move comfortably from model design into production reality. In the strongest conversations, the focus stayed on AI application architecture and the mechanics of training and inference, which suggests the team is looking for people who can reason about how models behave once they leave the notebook. Our candidate described those early discussions as thoughtful and relevant, and that’s a useful signal: eBay seems to value engineers who can connect ML choices to deployment constraints, not just explain algorithms in the abstract.
A recurring theme is how sharply the process can tilt toward low-level performance work. One candidate reported being pushed on GPU kernels, calibration, and quantization tradeoffs like GPTQ, AWQ, and GGUF, even after already covering broader optimization experience. That tells us eBay may care less about generic ML fluency and more about whether you’ve actually wrestled with inference efficiency and hardware-aware implementation details. The non-obvious risk here is scope mismatch: if your background is mostly model development or platform ML, you may still get pressed into very specialized territory.
We’ve also seen that ambiguity can become part of the evaluation. The coding exercise in this experience was described as unusually abstract, with unclear specs and a task that felt disconnected from the role. That means candidates should be ready not only for technical depth, but for situations where they have to impose structure quickly and make their assumptions explicit. At eBay, the signal seems to come from how you handle messy, performance-heavy problems when the prompt itself is not doing you any favors.
Synthetized from 1 candidates reports by our editorial team.
Had an interview recently?
Share your experience. Unlock the full guide.
Real interview reports from people who went through the Ebay process.
While I appreciated that eBay kept the process structured, the interview itself felt uneven in quality. The first two rounds were the strongest part for me: I spoke with a technical director about AI application architecture, and that conversation felt thoughtful and relevant to the role. The next discussion focused on model training and inference, and we spent most of the time on production implementation challenges, which was exactly the kind of ML engineering depth I expected going in.
The third round was where things went off track. It was a coding assessment, but the interviewer did not give clear technical specifications, and I was asked to write multiple parametric mathematical curves in C++. That felt disconnected from the actual scope of a machine learning performance engineering role, and the ambiguity made it hard to know what success looked like. The final technical round was also narrower than I expected. I explained my experience with model quantization, including GPTQ, AWQ, and GGUF tradeoffs, but the interviewer kept pushing on manual calibration techniques and then asked whether I had implemented GPU kernel fusion. That was a very specialized area and not something I had emphasized in my background. Overall, the process seemed to lean heavily into low-level GPU and kernel work rather than the broader training, data, and inference optimization work I associate with this role.
I did not receive an offer, but I appreciated the first two conversations and the chance to discuss AI infrastructure with strong technical people. My main takeaway is to be ready for very specific questions around GPU kernels, calibration, and performance optimization, and to expect at least one coding round where the problem statement may be unusually abstract.
Prep tip from this candidate
Be ready to discuss GPU kernel fusion and manual calibration in detail, not just higher-level ML performance work. Also practice handling an ambiguous C++ coding prompt, since the coding round used an abstract parametric-curve style problem rather than a standard ML coding exercise.
Share your own interview experience to unlock all reports, or subscribe for full access.
Sourced from candidate reports and verified by our team.
Topics based on recent interview experiences.
Featured question at Ebay
Explain what a p-value is to someone who is not technical
| Question | |
|---|---|
| Nearest Common Ancestor | |
| Rectangle Overlap | |
| Matrix Rotation | |
| Walking Robot | |
| Bias vs. Variance Tradeoff | |
| Target Indices | |
| String Palindromes | |
| Max Width | |
| The Longest Journey | |
| Seller Type Modeling | |
| Impossibly Iterative Fibonacci | |
| LRU Cache 1 | |
| Bias Variance Tradeoff | |
| Merge Sorted Lists | |
| Bagging vs Boosting | |
| First to Six | |
| Compute Deviation | |
| 500 Cards | |
| Permutation Palindrome | |
| Jars and Coins | |
| Hurdles In Data Projects | |
| Raining in Seattle | |
| Prime to N | |
| Get Top N Frequent Words | |
| Compute Variance | |
| Assumptions of Linear Regression | |
| Impression Reach | |
| Lazy Raters | |
| Biased Random Number Generator |
Synthesized from candidate reports. Individual experiences may vary.
The first round was a technical conversation with a technical director focused on AI application architecture. The discussion was thoughtful and role-relevant, with an emphasis on how ML systems are designed and integrated.
The next round dug into model training and inference, with most of the time spent on production implementation challenges. Expect questions around how you would build, deploy, and optimize ML workflows in a real engineering environment.
This round was a coding exercise with an unusually abstract problem statement. The candidate was asked to write multiple parametric mathematical curves in C++, and the lack of clear technical specifications made the expectations somewhat ambiguous.
The final technical round focused on low-level performance topics, including model quantization tradeoffs, manual calibration techniques, and whether the candidate had implemented GPU kernel fusion. The discussion was narrower and more specialized than the earlier rounds.