Google PhD New Grad AI/ML Onsite: Video Similarity, Expression Simplification, and Multi-Head Attention

Role: Machine Learning Engineer / SWE, PhD New Grad — NLP
Result: Passed the onsite and entered Team Match
Onsite date: March 17, 2025
Application: Applied online

Coding round 1: multiplicative video similarity

For a YouTube-style recommendation graph, similarity edges are weighted. If video A is similar to B and B to C, infer A to C with weight AB × BC. Return the k most similar videos.

The candidate modeled the data as a graph and used a greedy search. Follow-up: what if graph nodes are remote? They proposed replacing local attribute access with RPC calls.

Practice: Top K Videos by Multiplicative Similarity

The same round included a simple binary-tree task, but the candidate no longer remembered it.

Behavioral round

Questions included PhD research, handling difficulty, and other standard behavioral topics.

Coding round 3: simplify a nested variable expression

Simplify expressions such as a - (b + a - (c + a)) into combined coefficients, producing -b + a + c. The candidate used a stack for inherited signs and a counter for variable coefficients.

Practice: Simplify a Nested Variable Expression

ML/NLP round

The interview focused on transformers and LLMs:

• Factors affecting transformer inference latency
• Optimization techniques, tradeoffs, and complexity
• When to fine-tune and categories of fine-tuning
• Preventing private-data leakage in fine-tuning and production
• Collecting and processing structured versus unstructured chatbot data
• Implementing multi-head attention by hand during the final 15 minutes

Practice: Implement Multi-Head Attention

Google confirmed the pass on March 27. The candidate began Team Match in April and continued searching after two initial team conversations were not a strong fit.