Adobe · Primly Community

Adobe machine learning engineer interview: what the loop actually covers

ml_mike · 4 replies

Went through the Adobe machine learning engineer interview recently for a Sensei AI team role. "Sensei" is their ML platform brand. Sharing because MLE interview coverage for Adobe is sparse online.

The loop was six rounds total spread across two days for the virtual onsite.

Coding (2 rounds): Standard ML-adjacent coding. One round was essentially a data structures problem (nothing ML-specific, just algorithms), and one round asked me to implement a simple gradient descent loop in Python from scratch. No libraries, just numpy. They watched how I thought about numerical stability and step size.

ML fundamentals (1 round): Deep on theory. Bias-variance tradeoff, regularization, how you'd choose between L1 and L2, explain attention mechanisms at a conceptual level, when would you use a transformer vs. something simpler. They also asked about model calibration, which not everyone preps. Know Platt scaling at least in concept.

ML system design (1 round): Design a content recommendation system for Adobe Stock. Starting from raw event data through feature engineering, model selection, serving, and A/B testing. I spent a lot of time on the feature engineering and cold start problem because that's genuinely hard for a stock image use case. The interviewer seemed to like that I didn't jump straight to "just train a two-tower model" without justifying it.

Past project deep-dive (1 round): Very similar to the DE loop above: pick a project, they go three layers deep. What broke, what metric you picked and why, what you'd do differently.

Behavioral (1 round): Standard STAR. Adobe values cross-functional collaboration heavily in their behavioral questions. I got: tell me about working with PMs on an ML project where they wanted something you didn't think was technically feasible.

Model calibration and the cold start problem were the two things I'm glad I'd prepped. The gradient descent from scratch round would have gone badly if I hadn't done it at least once before walking in.

4 replies

ds_dmitri

The model calibration question is underrated. I've gotten it twice in 2025-2026 interviews and both times it felt like a real differentiator. Most candidates give a vague answer. Knowing that Platt scaling assumes sigmoid-shaped miscalibration vs. isotonic regression being more flexible is the level of depth that lands.

infra_ines

Six rounds is a lot. Was this all on the same day or split across two days? Also did they give you any heads up on what each round would cover or did you have to guess from the calendar invite titles?

ml_mike

Split across two days, three rounds each. The calendar invites had vague names like "Technical Discussion 1" and "Team Fit" so I just prepped everything. The recruiter did tell me there would be a system design round and a coding round, which helped narrow it somewhat.

returner_ren

Appreciate this. Coming back after a gap I've been worried about ML theory questions. Good to know calibration and gradient descent mechanics are worth reviewing. Those are the kinds of things that get rusty.