Amazon SDE VO Interview Experience (2026) - Jump Game II, Binary Tree Cameras, Remove K Digits, Offer

Problem 2: Binary Tree Cameras

Problem: Given a binary tree, install the minimum number of cameras so every node is monitored. A camera on a node monitors itself, its parent, and its immediate children.

Brute force was obvious and obviously bad: try every subset of nodes. The interviewer was waiting for me to move on. I shifted to post-order DFS with three states per node:

• 0: not covered
• 1: covered but does not have a camera
• 2: has a camera

The transition: if any child is in state 0, the current node must install a camera (state 2). If any child has a camera (state 2) and no child is state 0, current node is covered (state 1). Otherwise, current node is state 0 and its parent will need to install.

Null children return the "covered" state, which is critical so the leaves don't force their parent into camera installation. A final check at the end handles the root being uncovered.

Complexity: O(n) time, O(h) stack. The interviewer asked if I could prove the greedy bottom-up placement is optimal. I argued by exchange: placing a camera on a leaf instead of its parent loses coverage of the grandparent, so the parent placement strictly dominates. They accepted that.

Round 2: Coding plus Behavioral

Problem: Remove K Digits

Problem: Given a numeric string num and integer k , remove exactly k digits so the resulting string represents the smallest possible number. Return the result as a string.

My first instinct was "try all C(n, k) removal combinations," which I did not even say out loud because the combinatorial blowup is obvious. I went directly to the monotonic stack approach.

Key insight: we want the leftmost digits to be as small as possible. Walk the string and maintain a stack. For each digit, while the stack top is greater than the current digit and we still have removals budget, pop. Push the current digit. If k is still positive after the walk, remove from the tail (the remaining digits are in non-decreasing order, so the largest are at the end).

Two subtle bugs the interviewer probed for:

• Leading zeros. After the algorithm, you might have "0012" as the result. Strip leading zeros. If the string becomes empty, return "0".
• `k` not exhausted. If the input was already monotonically non-decreasing, the while loop inside the main pass never runs, and you need the tail-pop at the end.

I flagged both before coding, which saved me a follow-up.

Practice it: [[problem/159?company=3|Optimize Box IDs (Lexicographically Minimal String)]]

Behavioral: LP High-Frequency Questions

The second round bundled a 20-minute BQ block at the end. Questions:

• Tell me about a project you are proud of
• Describe a time you had to make a decision with incomplete data
• Tell me about a time you disagreed with a teammate

I had initially answered the "proud project" one with "I built an automated reporting system," which is exactly the kind of empty top-line statement that reads as padding. During prep I had rewritten it around the STAR frame: the PM pain point before the project, the specific technical choice I made (async pull with incremental cache vs sync pull per request), and the measured 30% drop in report generation time after launch. That version lands.

The "incomplete data" question maps to Bias for Action. I used an internship story about shipping a default config before full benchmarking because benchmarking would have delayed launch by 2 weeks, and the risk was bounded because we had easy rollback.

Round 3: Bar Raiser

This round was 90% behavioral. One short coding warm-up (two-sum-style check-in), then 45 minutes of LP deep-dive. The bar raiser has no domain overlap with your role, so don't lean on technical jargon.

Probe questions: most controversial technical decision you've made, when you last changed your mind after getting new data, and what your previous manager would call your biggest weakness. The "change your mind" question is the Bias for Action plus Dive Deep pair. The failure mode is saying "I never really had strong priors, so I'm flexible." That reads as lack of conviction.

Result

Recruiter called with the verbal offer 5 business days after the VO. SDE 1, Seattle, standard new grad comp. Written offer landed 3 days later.

Tips

1. Pivot fast when your first approach is O(n^2) and n is large. On Jump Game II I nearly started coding DP before catching myself. The interviewers do not penalize a wrong first instinct, but they penalize 10 minutes of writing a wrong solution. Call it out in 30 seconds.
2. Binary Tree Cameras needs a three-state DFS. Memorize the state definitions and the transition rules. The null-child returns 1 (covered) trick is the entire reason the algorithm works, and forgetting it will cascade into every test failing.
3. Remove K Digits is a monotonic stack test. If you see "remove k characters to minimize/maximize," your default tool is a monotonic stack. Always handle the leading-zero and unused-k tail cases explicitly.
4. Rewrite your "proud project" story. If it reads like a resume bullet, it is too shallow. Add the pain point, the technical decision, and a quantified outcome. Amazon interviewers literally score on whether you can quantify.
5. Prep a story for "changed my mind." This is the bar raiser's favorite question and is unusually hard to answer on the fly. Have one story where you had a confident technical prior, saw data, and revised.
6. Treat the bar raiser as pure LP. Do not try to impress with technical depth in this round. Your best signal is a coherent, self-aware story that maps cleanly to two or three LPs.

Amazon's full-time loop rewards candidates who are rehearsed but not robotic. Put real hours into LP story structure, drill your top 5 medium-hard algorithms until you can write them fast, and the loop stops feeling mysterious.