August 2026 Amazon Non-New-Grad Full-Time OA & Interviews: 23 Experiences

A practical roundup of 23 firsthand Amazon candidate reports.

Coverage: Non-New-Grad Full-time candidates only; SDE, SDE2, L4, L5, L6, Applied Scientist, and Engineering Manager; OA, Work Simulation, phone screen, virtual onsite, and team loops; reports published August 2–30, 2026.

For those full-time experiences with no specified roles, they are very likely to be SDE2.

The short version

SDE2 reports leaned especially hard on repository debugging, design ambiguity, or follow-up depth.

Leadership Principles mentioned in these reports

Only Learn and Be Curious and Deliver Results were named directly by a candidate. The other labels below are the closest LP mappings to the behavioral prompts candidates reported.

1. Customer Obsession: customer-experience metrics and feedback, plus a time the candidate rejected a customer request.
2. Ownership: taking responsibility beyond the immediate team, recovering from missed requirements, and improving a process.
3. Invent and Simplify: using GenAI to improve efficiency and simplifying or optimizing a process.
4. Are Right, A Lot: making difficult decisions, explaining trade-offs, and acting when metrics conflicted with customer feedback.
5. Learn and Be Curious: explicitly named; learning new approaches and reflecting on lessons learned.
6. Hire and Develop the Best: handling a low performer and motivating team members.
7. Insist on the Highest Standards: raising a quality bar and responding when a requirement was not met.
8. Think Big: building work that benefited other teams and reasoning about much larger scale.
9. Bias for Action: working through tight or impossible timelines and changing direction when needed.
10. Earn Trust: responding to criticism, handling conflict, and working through disagreement.
11. Dive Deep: finding root causes, handling detailed follow-ups, and reconciling data with customer feedback.
12. Have Backbone; Disagree and Commit: strongly supporting one direction, then working through the team's final decision.
13. Deliver Results: explicitly named; delivering under urgent deadlines and tying decisions to measurable outcomes.
14. Strive to be Earth’s Best Employer: balancing challenging work with team health.

Live interview question bank

Algorithms and stateful services

1. Track Package Receipt Gaps: longest contiguous prefix from zero plus missing IDs.
2. Arrange Packages With Compatible Adjacent Weights: decide feasibility in O(n) with a ten-slot frequency array; Rearrange Digits Under an Adjacent-Sum Limit: construct a valid digit order.
3. Count Repeat Customers Across Dates: filter to one time window and handle a very large log.
4. Streaming Top K Frequent Elements, Sliding Window Maximum, and LRU Cache for Query Results.

Design and OOD

1. Package Dependency Order: install dependencies first, return build order, and detect cycles.
2. Design a Seller Product-Image Upload Service.
3. Design a Marketplace Product Search Service.
4. Industrial code-based OOD that expected patterns/generics rather than simple conditionals.
5. Design an ML Training and Deployment Platform.

Science and management

Statistics, experiment design, anomaly detection, transformers and attention variants, Bootstrap Mean and Confidence Interval, and memory-versus-compute-bound MLP latency calculations all appeared in non-SDE loops.

23 detailed candidate timelines, loops, and screens

01: OA with commit-priority DP and Movie DB debugging (Ignore, it is about OA)

Report date: August 30 · Role: Full-time SDE · Result: WaitList

1. Minimum Merge Conflicts: merge two commit-priority strings while preserving both internal orders and minimizing inversions.
2. Debug a Spring Boot Movie DB CRUD service against six tests. The closest FastPrep AI Project is Repair MovieDB Search and Recommendations; it practices the same repository-repair family in a reviewed workspace. AI could suggest directions but could not edit the solution.
3. Candidate spent about 45 minutes and felt slowed by unfamiliarity with Spring Boot.

02 is about OA, skipped.

03: Experienced-hire phone screen with clock angle

Report date: August 26 · Role: Full-time SDE · Result: Pass

1. Recruiter highlighted Learn and Be Curious and Deliver Results.
2. Behavioral answer received a trade-off follow-up.
3. Distinct Duplicate Values.
4. Smaller Clock Hand Angle: candidate corrected an initially wrong direction through interviewer discussion, completed the whiteboard code, and dry-ran it.
5. Passing result arrived the next day.

04: OA with contiguous-quality inventory and API tests (Ignore, it is OA)

Report date: August 25 · Role: Full-time SDE · Result: Not disclosed

1. Minimum Contiguous Replacements: replace whole quality-value groups so every value forms one block using the fewest replacements.
2. Repair controller validation, persistence, and response-status tests in an existing repository.

05: L4 phone screen, sorting to frequency counting

Report date: August 24 · Role: L4 SDE · Result: Fail
Arrange Packages With Compatible Adjacent Weights: candidate first proposed sorting and alternating large/small values; the interviewer prompted an O(n) solution using the fact that values are only 0–9. Interviewed Friday and rejected Monday.

06: Alexa Shopping L6 loop

Report date: August 24 · Role: L6 · Result: Fail

1. Design a Seller Product-Image Upload Service.
2. Implement a Min-Priority Queue, matching the manual-heap requirement.
3. Design a Marketplace Product Search Service.
4. Streaming Top K Frequent Elements: count the most frequent IDs among millions of requests; candidate wrote pseudocode and complexity.
5. Behavioral topics included GenAI, impossible timelines, customer metrics, quality bars, hard decisions, and team disagreement.

07: Thin OA sighting—Inventory Allocation plus Django Issue creation (OA)

Report date: August 21 · Role: Full-time SDE · Result: Not disclosed
Coding was named Inventory Allocation by Bid Priority. Repair Workflow Issue and Sub-Issue Creation is the matching FastPrep AI Project for the Django/Python Issue-creation family. The source supports no deeper mechanics.

08: Repeat-customer logs at scale

Report date: August 19 · Role: Full-time experienced-hire SDE · Result: Pass

1. Behavioral: root-cause discovery and process optimization.
2. Count Repeat Customers Across Dates: count customers who visit on different dates.
3. Follow-ups: a unified time window and an extremely large log.
4. Passing notice arrived about two days later.

09: SDE2 Resilience Hub phone screen

Report date: August 19 · Role: SDE2 · Team: Resilience Hub · Result: Pass

1. Twenty minutes behavioral and 30 minutes DSA.
2. Behavioral: root-cause deep dive and tight-deadline delivery.
3. Rearrange Digits Under an Adjacent-Sum Limit: candidate clarified requirements, used greedy reasoning, handled edge cases, proposed tests, dry-ran two, and discussed complexity.
4. Next-round email arrived the following day.

10: Hard OA with grid BFS and Movie DB debugging (OA)

Report date: August 18 · Role: Full-time experienced hire · Result: Candidate expected failure

1. Minimum Grid Inconvenience, described as multi-source BFS plus binary search.
2. Repair MovieDB Search and Recommendations is the closest FastPrep AI Project; this report used Spring Boot with six failing tests, and the candidate completed only the first two test groups before running out of time on persistence/status issues.
3. Python was another available framework choice.

11 is about NG interview experience, so I deleted it.

12: Stateful package receipt tracking

Report date: August 14 · Role: Full-time experienced-hire SDE · Result: Pass
Track Package Receipt Gaps: accept out-of-order package IDs and report the contiguous prefix endpoint from zero plus missing IDs. Example: 0,1,2,3,100 returns 3 and 4..99.

13: SDE2 loop with industrial OOD

Report date: August 14 · Role: SDE2 · Result: Fail

1. Four rounds; all contained Leadership Principles.
2. Coding was unfamiliar but array/hash-map level.
3. System design began as production architecture before the interviewer clarified infrastructure intent.
4. OOD used dense industrial Java-like code, unexplained object fields, and an expected pattern/generics solution.
5. Rejected after more than a week and told the cooldown was one year.

14: L6 Applied Scientist breadth screen

Report date: August 14 · Role: Applied Scientist L6 · Team: Amazon Ads · Result: Pass

1. Thirty-minute rapid-fire technical screen with no behavioral section.
2. Statistics, A/B testing, experiment design, power analysis, anomaly detection, bagging/boosting, transformers, attention variants, and model families.
3. Bootstrap Mean and Confidence Interval, plus optimization without a loop.
4. Passing notice arrived two days later.

15: Exact distribution-center OA and Issue debugging (OA)

Report date: August 14 · Role: Full-time SDE · Result: Not disclosed

1. Minimum Grid Inconvenience: binary grid, Chebyshev distance, n,m <= 500, add one center, minimize inconvenience.
2. Candidate used BFS, feasibility checking, and binary search.
3. Repair Workflow Issue and Sub-Issue Creation: request validation, ID generation, Issue creation, and schema compliance.

16: SDE2 flash-sale inventory and Movie Search (OA)

Report date: August 11 · Role: SDE2 · Result: Not disclosed

1. Unfulfilled Customers by Inventory Priority: allocate stock by bid, round-robin tied bids in timestamp order, and return customers receiving nothing.
2. Repair MovieDB Search and Recommendations: this report used Spring Boot, where one defect failed to lowercase a partial-title query while stored values were lowercase.

17: Engineering Manager ML-systems calculation

Report date: August 8 · Role: Engineering Manager · Result: Not disclosed

1. Handle a low performer; balance challenging work with team health; motivate a team.
2. Given an MLP, HBM 1 TB/s, SRAM 25 MB, BF16 100 TFLOPS/s, and dimension 4096, calculate batch-1 and batch-256 latency and prove which is memory- versus compute-bound.

18: Greedy inventory plus Django/React debugging

Report date: August 7 · Role: Full-time SDE · Result: Fail

1. Get Minimum Amount: make each quality value contiguous at minimum replacement cost; greedy interval-merge/sweep-line family.
2. Repair MovieDB Search and Recommendations: debug unfamiliar Django + React code from unit tests.
3. Java, C++, and Python were offered.

19: L5 boomerang, two team outcomes

Report date: August 6 · Role: L5 boomerang · Result: Offer

1. First team: Print a Shortest Graph Path, weighted lottery selecting k winners, and Design an ML Training and Deployment Platform.
2. Candidate felt the ML design mismatched the role.
3. Second team: whiteboard the candidate's own architecture and solve an Amazon Locker problem.
4. Offer arrived two hours later because another deadline was imminent.

20: Senior AWS package tracker

Report date: August 5 · Role: Senior SDE · Location: Bay Area · Result: Fail
Behavioral covered an urgent timeline, rejecting a customer request, and lessons learned. Coding repeated Track Package Receipt Gaps; another candidate independently said a hash map plus one variable solved it.

21: Dependency installer and build order

Report date: August 4 · Role: Full-time experienced-hire SDE · Result: Not disclosed

1. Phone: implement installWithDependencies, install dependencies first, and detect cycles. Package Dependency Order is the matching FastPrep practice version.
2. Onsite: return build order for target package x using DFS topological sort, Visiting/Visited, and O(V+E) analysis.

22: SDE2 OA-to-phone pipeline

Report date: August 4 · Role: SDE2 · Result: Advanced to phone screen

1. OA taken August 3: 100-minute algorithm plus AI-repository section.
2. Unfulfilled Customers by Inventory Priority, plus repository debugging whose source did not preserve enough detail for a separate practice project.
3. Work Simulation about 15 minutes; two Work Style surveys about 10 minutes.
4. Phone invitation August 4; scheduled for August 14 at 3:00 p.m. Pacific.

23: SDE2 maximum value from k consecutive bags

Report date: August 2 · Role: SDE2 · Result: WaitList
Find Maximum Total Amount practices the same non-overlapping [l,r,value] intervals and exactly-k window core; the active FastPrep canonical also includes a modulo return rule from richer evidence.