Netflix — Machine Learning Engineer ❌ Failed

Level: Unknown Level

Round: Phone Screen · Type: Coding · Difficulty: 7/10 · Duration: 60 min · Interviewer: Unfriendly

Topics: Machine Learning, Transformers, BERT, Tokenization, Data Structures, Algorithms

Location: Los Gatos, CA

Interview date: 2025-05-15

Got offer: False

Summary

I interviewed with a recruiter for a Machine Learning Engineer position. The technical phone screen was with an interviewer who asked many questions about my projects for the first 25 minutes. After that, the interviewer asked transformer-related questions.

Details

I was asked about sentence piece, what tokenizers BERT and Transformer use, the components of a Transformer block, and the architectural differences from Transformer to Llama/Qwen. I was also asked about the benefits of these differences (MoE, RMSNorm, rotary positional encoder).

The coding question I got was to convert a Transformer dictionary from a list of layer1.attention… format to a dictionary format. I needed to put similar layers in the same layer.

`python

Example Input (Conceptual)

[ "layer1.attention.weight", "layer1.attention.bias", "layer2.attention.weight", "layer2.attention.bias", "layer1.ffn.weight", "layer2.ffn.bias" ]

Expected Output (Conceptual)

{ "layer1": { "attention": { "weight": ..., "bias": ... }, "ffn": { "weight": ... } }, "layer2": { "attention": { "weight": ..., "bias": ... }, "ffn": { "weight": ... } } } `

My approach:

1. First, I split each string by the "." delimiter.
2. Then, I recursively constructed the dictionary, creating nested dictionaries for each level of the split string (layer, module, parameter).
3. I iterated through the list and populated the dictionary with the corresponding weights or biases.