Amazon Applied Scientist Interview Experience: Alexa Speech Team, 2021

How the process went

1. Recruiter outreach

   An Amazon recruiter first called about an Amazon Go team opening; the candidate expressed interest in Alexa and other NLP teams instead, which led to a phone screen with the Alexa Speech team.

2. Phone screen

   A roughly one-hour phone screen in July mixing a leadership-principles question, an end-to-end ML design discussion, and a short coding question in the last five to ten minutes.

3. Scheduling

   An onsite invitation followed within a couple of days; the candidate initially scheduled it two weeks out, then pushed it back about a week further for personal reasons.

4. Virtual onsite

   A split loop across two interested teams: a technical presentation to a cross-team audience, then five rounds of an hour or more each, including two hiring-manager rounds, a bar raiser, and two peer panels, each covering a different set of leadership principles.

5. Additional round

   About two days after the main loop, the recruiter scheduled one more round with a senior applied scientist specifically to check ML breadth further.

6. Offer and leveling

   Around two weeks after the additional round, the recruiter confirmed an offer. Team matching then took a while, and the candidate was downleveled from L5 to L4 due to limited production-environment experience.

Phone Screen — Alexa Speech Team

Leadership principles, an end-to-end ML design discussion, and a short coding warm-up · About 1 hour, with the last 5-10 minutes for coding

• A leadership-principles discussion centered on bias for action.
• An end-to-end ML design question: given a problem statement, walk through problem formulation, data gathering, data processing, model choice, evaluation metrics, and how to explain those metrics to non-technical stakeholders.
• Follow-up questions challenging specific choices along the way, such as why one model or metric was chosen over another.
• A short, easy tree coding question solvable with a one-line recursive solution (the candidate did not recall the exact prompt).

Technical Presentation

Research talk to a cross-team audience

• Presented an in-depth talk on one of the candidate's own research projects to attendees from multiple interested teams, followed by a few clarifying questions about specific slides.

Hiring Manager Round — Team 1

Team fit plus an end-to-end ML design question tied to the team's actual problem

• Standard leadership-principle questions with follow-ups.
• An end-to-end ML approach for the team's own problem area, covering data issues, model choice, and metrics.

The hiring manager said the candidate's approach matched how the team was already tackling the problem.

Hiring Manager Round — Team 2

Leadership principles, NLP breadth, and a coding exercise

• A different set of leadership-principle questions than the earlier hiring-manager round.
• A general discussion of state-of-the-art NLP models and when to use them.
• A coding problem similar to Word Break I and II, solved with backtracking and walked through with an example.

The interviewer was an alum of the candidate's then-current research company, and the round included a reverse-interview segment on the team's work and expectations.

Bar Raiser Round

Heavy leadership-principle probing plus ML metrics discussion

• Multiple in-depth leadership-principle questions with detailed follow-ups on past decisions and pivotal moments.
• Questions on which ML evaluation metric to use and why, explained at a non-technical level.

The interviewer was a senior business-development lead from the AWS org with over 20 years of experience.

Team 2 Panel — ML Breadth and Depth

Broad ML fundamentals with deeper follow-ups where answers went further

• Questions on regularization, loss functions, and model architectures.
• Questions on speech recognition and language modeling.
• A deeper discussion of normalization techniques for regularization and training speed-up when the candidate's answers went further into detail.

Two interviewers from team 2 with similar experience levels.

Team 1 Panel — ML Breadth

Rapid architecture questions

• Rapid-fire ML breadth questions on LSTM and other recurrent-architecture equations, plus CNNs.

One of the two interviewers was shadowing. The candidate felt a communication gap with the primary interviewer, who needed the shadow interviewer to step in and clarify some answers.

Additional Round — Senior Applied Scientist

A second end-to-end ML design question to probe breadth further

• Given a problem scenario, design an end-to-end ML solution covering problem formulation, scope reduction, data gathering, domain adaptation, knowledge distillation, language models, and metrics.

Added specifically to check ML breadth after the main loop. The candidate narrowed the problem toward more familiar territory during the discussion, solving smaller pieces before offering options to expand.