A practical roundup of five firsthand Amazon SDE II/L5 reports published after September 20, refreshed through October 11, 2026.

Coverage: Reports published September 22–October 9, 2026; Amazon SDE II / L5 phone screens and onsite loops; coding, LLD, HLD, Bar Raiser, Leadership Principles, AI questions, and project deep dives. Locations include Canada and Bengaluru, while two candidates did not state a location. The date window follows report date, not interview date: one loop happened in May–June, and the September 22 post describes a September 16 phone screen.

The short version

Across these five reports, Amazon SDE II looked less like a list of isolated coding questions and more like a test of how candidates handled changing rules, incomplete requirements, design depth, and repeated evidence checks.

1. Design remained central in the complete loops. Candidates reported LLD, HLD, a dependent-job scheduler, a delivery-estimation domain, and one LLD round that required runnable code.
2. Coding answers kept expanding. Interviewers asked candidates to preserve directional state, adapt chess movement rules, move from recursion to memoization, print a heap merge path, and defend time-versus-space choices.
3. Leadership Principles stayed inside technical rounds. One candidate reported two LP questions in every round; another spent about twenty minutes on resume depth, tradeoffs, personal contribution, and two LP questions before coding.
4. AI appeared as a behavioral topic. One L5 screen asked how the candidate verifies AI output, creates team benefit with AI, and investigates incorrect AI results.
5. The process shape varied. The reports include a phone screen, four-round loops, and a five-round sequence. Two reports disclose rejection, one says no offer, and two do not provide a final outcome.

A practical preparation formula: narrated DSA + explicit state invariants + executable LLD + one failure-aware HLD + quantified LP stories.

Behavioral questions and project deep dives

The reports preserve these concrete questions and follow-up directions:

1. How do you verify output produced by AI?
2. How have you used AI to create a real benefit for your team?
3. When an AI result is wrong, how do you investigate it?
4. Why did you choose this technology or approach, and which tradeoffs mattered?
5. What business rule did the project implement, and what did you personally own?
6. What measurable result followed—scale, latency, reliability, cost, adoption, or another concrete outcome?

The September 22 phone screen spent roughly twenty minutes on the candidate’s resume, technology choices, tradeoffs, business logic, personal contribution, and two LP questions. The Canada candidate reported two LP questions in every round, with follow-ups seeking metrics and measurable impact. The Alexa Devices report gives behavioral-question counts by round but not the wording.

Practical BQ preparation: for every story, write down the starting state, your personal decision, one rejected alternative, the hardest follow-up, a before-and-after metric, and what you would change now.

Live coding question bank

Graphs, trees, tries, and dynamic programming

1. Grid Reachability with a Forbidden Turn. A grid contains open cells, blocked cells, a start, and a destination. Movement depends on the current heading; reversal is disallowed and one turn direction is forbidden. The candidate used BFS but later noticed that the visited state did not preserve direction.
2. Minimum Moves to Capture a King. Find the shortest chess path to the king, then handle a follow-up in which allowed movement directions depend on the piece type. The report does not preserve every variant rule.
3. Concatenated Words variation. The candidate described a LeetCode 472-style task and explained a trie/DFS approach aloud. The exact variant details were not disclosed.
4. Binary-tree zigzag variation. The candidate recognized spiral level order, but the change from the standard contract was withheld. Use this related spiral level-order practice as pattern work, not an exact reconstruction.
5. 0/1 Knapsack variation. The candidate moved from recursion to top-down memoization and discussed complexity. Weights, capacity, return value, and the actual variation were not published, so no exact problem is linked.
6. Minimum cost to connect ropes. A long Alexa-device story hid the min-heap pattern. The follow-up required recording and printing the exact merge path, which the linked practice problem does not yet cover.
7. Two Canada coding rounds. One undisclosed Medium required brute force, optimization, and a time-versus-space discussion. A second problem arrived in three incremental parts and required a long implementation. No exact links are attached because both prompts were withheld.

LLD and HLD prompts

1. Delivery-estimation domain. Discover functional and non-functional requirements, then turn them into a structured object model. The candidate believed weak requirement discovery was the main reason the packet failed.
2. Distributed recurring workflow scheduler. Discuss dependency ordering, queues and partitioning, storage, retries with backoff, horizontal scaling, monitoring, and logs.
3. Runnable LLD in Canada. The domain was withheld, but the candidate emphasized that a class diagram or discussion was not enough; working code was expected.

Five detailed candidate timelines

01: Bengaluru five-round sequence

Report date: October 9, 2026 · Role: SDE-2 · Location: Bengaluru · Result: No offer

The live process contained two DSA rounds, LLD, HLD, and a Bar Raiser. No prompt was published. These coding, LLD, and system-design libraries fit the round types; none is claimed as an exact match.

02: Alexa Devices loop where LLD decided the packet

Report date: October 7, 2026 · Interview timeline: May–June 2026 · Role: SDE II · Team: Alexa Devices · Location: Bengaluru · Result: Rejected

1. DSA: a binary-tree zigzag variation (related spiral level-order practice); an unlinked 0/1 Knapsack variation whose actual contract was withheld; recursion to memoization; complexity; two LP questions.
2. LLD: delivery-estimation service, with requirement discovery before design; two LP questions.
3. HLD: dependent-job scheduler covering ordering, queues, storage, retries, scale, and monitoring; three behavioral questions.
4. Bar Raiser: connect ropes with a merge-path follow-up; three behavioral questions.

The candidate felt coding and HLD went reasonably well but believed LLD was the main weakness.

03: L5 phone screen with chess, trie, and AI questions

Report date: October 7, 2026 · Role: L5 SDE / SDE II · Status: Awaiting result when posted

1. Solve Minimum Moves to Capture a King, then adapt movement rules by piece type.
2. Explain a trie/DFS solution for a Concatenated Words variation.
3. Discuss how to verify AI output, create team benefit with AI, and investigate incorrect AI results.

The report preserves the patterns but not the complete variant contracts.

04: Canada L5 loop with runnable LLD

Report date: September 24, 2026 · Role: SDE II / L5 · Location: Canada

1. Coding one: move from brute force to an optimized solution and defend time-versus-space choices.
2. Coding two: solve a three-part incremental problem with lengthy implementation.
3. LLD: produce running code for an undisclosed domain.
4. HLD: continue verbally in CoderPad after the design-board link was missing.

Every round contained two LP questions with follow-ups about metrics and measurable impact. All four prompts were withheld, so the coding, LLD, and system-design links are broad libraries, not exact matches.

05: September phone screen with a direction-state bug

Interview date: September 16, 2026 · Report date: September 22, 2026 · Role: SDE II · Result: Rejected about one week later

The first twenty minutes covered the resume, technology choices, tradeoffs, business logic, personal contribution, and two LP questions. Coding then used Grid Reachability with a Forbidden Turn. The candidate found the BFS direction-state bug after the interview. Replies clarify that one turn direction is forbidden; they do not prove that both turns are blocked.

What repeated often enough to prioritize

Tier 1: Practice until automatic

1. Solve while explaining. Move from requirements to baseline, optimization, complexity, invariants, edge cases, and a dry run without going silent.
2. Turn requirements into code. The strongest design warning in this five-report set is that SDE II LLD may need executable behavior, not only boxes and arrows.
3. Prepare behavioral evidence with numbers. State what you personally changed and what measurable result followed.

Tier 2: Be ready to explain deeply

1. State and follow-ups. Direction, recursion state, heap operations, and incremental requirements all changed what counted as a complete answer.
2. Failure-aware design. For a scheduler, discuss persistence, duplicate work, retries, backoff, partitioning, recovery, and observability.
3. Project depth. Explain technology choices, business rules, rejected alternatives, and your personal contribution without relying on a framework list.

Tier 3: Current-cycle differentiators

1. AI judgment. One current L5 report explicitly asked about verifying AI output, team benefit, and investigating wrong results.
2. Adaptation under ambiguity. The reports include missing design tools, incremental prompts, and partially disclosed task contracts. Clarify what is expected, state assumptions, and keep moving.

Final checklist

Coding

☐ Practice directional BFS, tree traversal, tries, memoization, heaps, and incremental requirements.

☐ After the first solution, test changed constraints, state invariants, edge cases, and complexity.

LLD and HLD

☐ Timebox requirement discovery and start the executable happy path early enough to finish core methods.

☐ For a scheduler or service, discuss scale, persistence, ordering, retries, idempotency, recovery, and observability.

☐ Separate confirmed requirements from practice assumptions and from your own interview assumptions.

Leadership Principles

☐ Prepare real stories for ownership, tradeoffs, customer impact, mistakes, disagreement, and measurable results.

☐ For AI stories, prepare verification, safeguards, adoption, manager reaction, and quantified impact.

☐ Keep your timeline, personal contribution, rejected alternatives, and metrics consistent through repeated follow-ups.