What candidates say the round looked like, summarised in our words, each linked to where it came from. Amazon hasn’t confirmed these details unless the source says “Official page”.
Background on the format: our Amazon guide.
Had this round? Add your report →
Finished the algorithm question, then in the AI round, done in C++, figured out where the issue was and wrote a fix, but none of the test cases passed even after running them repeatedly. Knowing the bug isn’t enough; the fix has to satisfy the tests.
Intern assessment: the algorithm question was easy to medium and solvable with a greedy approach; the AI-assisted task was medium. Read the question and run the 6 test cases. Paste a failing test into the AI and it tells you where to look and which function to debug; then go through that function carefully yourself.
The AI-assisted task is straightforward if you know the framework you pick (Django, Spring Boot or Node; this candidate picked Spring): check which tests fail and why. Using the AI wasn’t required to pass; they asked it for maybe two changes, mostly to debug an existing function. Don’t try to understand all the code, just clear the tests. Many bugs are trivial, like a missing slash in a URL or a missing null check.
The new online assessment pairs a traditional algorithm problem with a live AI debugging question. You can use the AI to understand what the question is asking, but the debugging and the fix are yours to do.
Online assessment with one algorithm question and one AI-assisted coding task: a full web app in Node.js and React to fix. The AI won’t fix it for you. It only helps you find the bugs or walk through the project. The candidate found the app task tougher than the algorithm question.
SDE-1 online assessment in the one algorithm question plus one AI-assisted repo question format. The candidate found the repo question very easy. The algorithm question, a greedy bit manipulation problem, was the part they couldn’t solve, so the repo task was not where this assessment got hard for them.
US SDE 2026 assessment: the traditional question was solved with every test passing, but on the multi-file repo with the AI assistant the candidate struggled to isolate the bug across the codebase in time and passed only 1 test.
SDE-1 assessment with the opposite split to most reports: the candidate couldn’t finish the algorithm problem (0 cases passed) but passed all 6 of 6 tests on the AI-assisted development task, and was waiting to see if that was enough for an interview.
6 of 15 on a greedy algorithm problem, then half an hour just getting familiar with the codebase and learning how to get answers out of the AI, which left time for only one bug. They didn’t know Spring Boot well. The AI won’t hand you the fix, but questions like ‘what does this method do?’, ‘why is this test failing?’ and ‘which classes should I look at for this bug?’ save a lot of searching.
Intern online assessment on HackerRank. The first question was easy, solved in 15 minutes with 15 of 15. The AI repo question was much harder because there were so many files to navigate, and they got 2 of 6 tests passing.
SDE-2 online assessment: the candidate passed every case on the algorithm question but couldn’t solve the AI assistant question. They had chosen C++, and the repo task also came in C++, with low-level code that was hard to understand in 60 minutes. No interview call followed. They planned to choose Java next time.
Short answer from someone who had taken it: the AI-assisted section asks you to debug one feature in a backend service, using the LLM available inside the IDE sandbox, until every test tied to that feature passes.
On-campus intern online assessment: one algorithm question (passed 15 of 15) and an AI-assisted repo debugging task. The repo was a support-desk style ticketing app where users open tickets and others reply in a thread. The failing behaviour: new tickets weren’t being saved, and replies to a ticket didn’t work, across 6 test cases. The candidate used Django and passed 2 of 6, yet was still shortlisted for interviews.
Intern assessment advice from someone who took it: don’t write new methods for the repo question; small changes to existing code are usually enough to make the tests pass. The assistant wasn’t much help; running the tests to see what they expect and where they fail was. The language you choose at the start is locked for the repo question.
Intern assessment: a medium-hard algorithm problem in 40 minutes, then 60 minutes on a repo question in the language you pick, such as a comment system, with 6 tests to make pass, for example a comment with no message returning 401. The AI doesn’t give direct answers and mainly helps with syntax and flow. A second candidate’s tip: run the tests and paste the errors into the agent window.
The algorithm question, about choosing consecutive bags from money intervals, hit time limits (4 or 5 of 15). The second part was an AI-assisted Java Spring Boot repo focused on role-based access control and security, and the candidate passed only 1 of 6 tests.
Intern 2027 assessment: 40 minutes for the algorithm question (11 of 15; 4 hidden cases failed, perhaps time limits) and 60 minutes for the AI code repository task (6 of 6, in Node.js). Their tips: do Amazon’s demo test, choose your language carefully, focus on what the tests check, and use the AI wisely.
Second question was AI-assisted C++ debugging: an existing C++ project with several source files plus JSON and JSONL config and data files. You understand the code, find the bugs and change it to produce the expected output. AI could analyse and explain the code and help debug, but not generate the solution. It was the candidate’s first task like this and they couldn’t finish it, after passing 15 of 15 on the first question.
SDE2 assessment in Dublin. The AI repository task was the commonly reported movie app where users follow actors and get notified about new movies. A few unit tests fail and you fix them; the AI helps with basic questions and exploring the codebase. Most of the time goes into exploring the controllers, services and tests. The candidate was surprised that about 70% of the work simulation was technical questions on reliability, scalability and storage choices.
Passed every test on the algorithm question and 6 of 6 on the AI coding task, yet heard nothing, and the application later switched to ‘no longer under consideration’. A perfect assessment score is not a guaranteed interview.
Assessment sent for SDE roles to someone with Applied Scientist and ML Engineer applications: 15 of 15 on a commit-history merging problem they’d seen in earlier posts, and 6 of 6 on the AI-assisted web development task in Django. The provided AI agent wasn’t much use to them: it didn’t point to the actual issue, so they mostly worked it out themselves.
Intern 2027 online assessment with two questions: a medium array problem, where an optimisation took the candidate from 12 to 15 of 15 tests, and a development task with an AI assistant built into the coding environment. On the development task they passed 2 of 6 test cases.
A dissenting view: this candidate found the assistant useless beyond syntax help and felt it wasted their time. Their advice is not to expect any major help from it and to plan on finishing the task yourself.
US SDE-1 online assessment: 15 of 15 on the algorithm question and 6 of 6 on the AI debugging task, done in Spring Boot. A week later Amazon emailed asking for some details, then nothing for two more weeks and no recruiter contact, so a perfect score didn’t mean a quick interview.
12 of 15 on the algorithm question, where 3 cases failed because the tests passed long values into HackerRank’s prewritten main(), and 1 of 6 on the AI debugging task. The odd part: each test passed when run on its own, but only 1 passed when the whole suite ran together. A week later the recruiter wrote that team matching had started.
AWS SDE2 online assessment in July. The AI-assisted section was a practical backend engineering question, and the AI was there so you could look up documentation without leaving for Google. Nobody expected vibe coding or AI planning: the assistant was barred from implementing a solution or writing code in the IDE.
SDE II in Redmond with about 6 years of experience, ending in an offer 51 days after applying. The assessment had one standard algorithm problem, one Spring Boot code refactoring problem with AI assistance, and the behavioural questions. A phone screen a week later and a four-round virtual loop followed.
The AI coding task was easy for this candidate: they solved it in 30 to 40 minutes because they already work with Node.js and could guess the problems in the controllers and routes. Their advice is to do the practice test Amazon provides, because the real question is similar.
A North American candidate who sat the assessment the day before also found the AI assistant in the second coding part not working, even though it had worked fine in the practice assessment. If it fails for you, note it, since others have hit the same problem.
SDE I online assessment in Europe. The algorithm question was a hard dynamic programming problem (3 of 15 passed). The AI-assisted task was a Spring Boot repo with 6 bugs, and the HackerRank AI assistant didn’t work at all: every message returned an error, even after reloading and switching browsers. The candidate worked through it without AI.
6 of 6 on the AI coding task. Their method: read the README about the project and its API endpoints, then the task (make one endpoint work), run the tests to see what fails, read the tests to see what they call, ask the AI to map the flow of what calls what for that endpoint, then fix one test at a time. Every test has to pass. They said it felt like being at work.
SDE online assessment: the candidate fully solved the traditional coding question (a Fenwick tree problem) and finished the work simulation and surveys, but passed only 2 of 6 test cases on the AI-assisted coding question before time ran out. They were waiting to learn how much weight the AI-assisted part carries next to the algorithm question.
SDE-1 online assessment: one medium algorithm question and one AI-assisted repo task, plus work simulation and survey sections. The repo ran in a remote IDE and the stack was the candidate’s choice of MERN, Django or Spring Boot. There were 6 test cases, meaning 6 bugs to fix in an hour, all in the backend: 4 in the controller, 1 in the routes and 1 in the middleware. The practice test had only 2 bugs. They passed every test with about 15 minutes to spare.
Answering a question about the SDE-2 online assessment: the AI-assisted part gives you a project repo with a frontend and a backend and a few issues to resolve so that 6 test cases pass. This candidate chose Node.js before the assessment started and had to fix a POST API in the repo.
India assessment: a medium dynamic programming problem (15 of 15) and a repo debugging task that felt like it was designed for SDE-2. The app was a movie database, with the notification and follow/unfollow features to fix, and working out what needed fixing took over 20 minutes. The AI won’t write code but explains errors and points you to the files and line numbers that need changes.
Node, Express and MongoDB backend for a review site with moderation and rate limiting rules to fix across controllers, middleware and utilities. The candidate skipped the tutorial and assumed the AI could edit code, which it can’t. They solved the first 3 or 4 issues. The AI helped with syntax but was of limited use for that many bugs, so they used print statements and tests to map which component caused which error code. Five minutes went to misreading the shape of an exported data structure.
SDE-II online assessment: one coding question and one issue-resolution task in a codebase with the AI assistant. The candidate fixed every issue in the acceptance criteria and all related tests passed. They then spotted another possible bug and started on it, leaving a half-written log line in the code at submission, and worried it could break the build. Lesson: stop once the acceptance criteria pass, or clean up before you submit.
This candidate passed the online assessment on their first try even though they passed only about half the test cases on the algorithm question. They believe the second question, the Django project with the AI assistant, carried a lot of weight, and that Amazon no longer requires a perfect algorithm score to shortlist.
This candidate ran out of time on the AI-assisted project. They understood the problem and implemented some of the functionality, but no tests passed. The biggest problem was practical: on a small laptop screen, switching between the file tree, the editor and the AI panel ate most of the hour. They didn’t expect an interview call.
SDE-1 online assessment in the new format: a 40-minute medium-to-hard algorithm question, then a 60-minute AI assistant round. The candidate got a codebase plus a preview of a website with three bugs to fix, with help from the assistant built into the assessment. Backend choices were Node.js with MERN, Java with Spring Boot or Python with Django. They thought it was fairly easy for anyone who already works with one of those frameworks.
The real task matched the practice one Amazon sends, except there were 6 bugs instead of 1. That was hard to fit in 60 minutes, and they solved only 2. They didn’t use the AI much and suggest practising with it, for example asking it to explain functions, to track bugs down faster.
SDE-2 online assessment with two timed coding questions: a 40-minute algorithm problem with no AI, then 60 minutes of software development with an AI assistant available. Before starting, the candidate picked a framework from Django, Spring Boot, C++ and Node.js and chose Django. In the repo, 6 unit tests were failing and the task was to fix the related APIs. Work style and behavioural sections followed.
SDE-2 assessment: a greedy array problem solved in 10 minutes, then repo debugging in a VS Code-like editor inside HackerRank, with an hour to make the unit tests pass in a medium-sized repo with frontend and backend. The candidate had never written frontend code, found a first bug there, and spent 45 minutes combing the frontend even after the AI said the bugs weren’t there, because they didn’t trust it. The backend had many bugs and they passed 2 of 5 tests.
The repo question was a wallet app, very similar in style to HackerRank’s practice repo: backend API bug fixes in paths, URLs and names. In total they edited about 10 lines, wrote 15 new ones and passed all 6 tests. The built-in AI gives you everything except the solution if you prompt it well with context, and it guided them to the right file and function. About 25 minutes on the algorithm question and 40 on the repo.
SDE II loop: the first round combined system design (an examination system) with SOLID and a leadership principle, plus one direct question: how do you use generative AI in your engineering workflow? Later rounds were algorithms, the bar raiser and a hiring manager design round. The candidate didn’t clear it.
SDE-1 third round, run by the hiring manager and labelled AI fluency and problem solving. The GenAI part was spoken questions: how you use GenAI in your work, how you leverage it, and what you learned from it. Then came algorithm questions; this candidate got two because they finished the first in 15 minutes.
SDE-1 loop: round 3 opened with a discussion of how the candidate uses AI tools while coding and which workflows they have automated, with follow-ups on implementation choices and practical use. The rest of the round was three algorithm problems plus leadership principle questions. No AI tool was used during the coding itself.
Intern loop where the second round mixed GenAI questions with algorithms. The AI part was a conversation: how much do you trust AI, what do you use it for, have you solved problems with it, which tools do you use. Then came two algorithm problems; the candidate had the logic for the second but couldn’t finish coding it in time and was rejected.
On-campus intern loop with a round titled GenAI Fluency. Despite the name, most of it was a heap-based algorithm question with brute force and optimal approaches, a dry run and a follow-up on other data structures. Only at the end did the interviewer ask which of the candidate’s projects used AI, how it was built, and a few basic GenAI questions. The candidate was selected.