Google PhD Early Career AI/ML: Triple Packing, Data Chunks, and Square Counting

Role: Machine Learning Engineer / SWE, PhD Early Career
Result: Passed an additional round and entered Team Match
Report time: March 2026

Timeline

• R1 virtual interviews: October 2025
• R2 onsite: December 2025
• Additional virtual coding round: January 2026
• Team Match notification: about two weeks after the additional round

R1 coding: stream triples by size

Items arrive on a belt. Store their sizes, and whenever three items satisfy
max(A, B, C) - min(A, B, C) < threshold, return that triple.

For one-dimensional sizes, the candidate inserted each value into sorted order with binary search and checked nearby consecutive triples.

A later follow-up changed each item to a two-dimensional point such as [3, 4] and required Euclidean distance. The candidate completed only part of that version and later suggested sorting by one dimension, filtering candidates, and evaluating remaining combinations. The exact two-dimensional contract was not preserved.

Practice: Streaming Triple Packing by Size Threshold

R1 behavioral

The behavioral round used standard questions and began with about 15 minutes of conversation.

R2 domain interview

The candidate presented a past project and discussed how it might be applied to Google products.

R2 coding: minimum data chunks

Split a data payload into as few chunks as possible, with every chunk at most a maximum size.

Example: data = 1800, maxSize = 700 returns balanced chunks [600, 600, 600].

The follow-up added an indivisible header. The header cannot be split across chunks.

Example: header = 600, data = 300, maxSize = 700 returns [600, 300].

The candidate did not reach the third follow-up because too much time had been spent discussing the earlier parts.

Practice: Minimum Data Chunks With an Indivisible Header

Additional coding round

The additional round used a BFS-style variation related to LeetCode 221 and asked for the total number of all-one square submatrices.

The candidate completed the problem but accidentally introduced an expensive sort and did not recognize the resulting complexity during analysis.

Practice: Count All-One Squares