← Back to experiences
[ OK ] Loaded —
[ INFO ]
$ cd
$ ls -lt
01
02
03
04
05
$ ls -lt
01
02
03
04
05
user@intervues:~/experiences/…$
Level: Senior-Level
Round: Full Journey · Type: Multiple Types · Difficulty: 7/10 · Duration: 360 min · Interviewer: Unfriendly
Topics: Coding, Tensor Programming, Schema Validation, Behavioral Questions, System Design, Distributed Systems, Machine Learning, Deep Learning
Location: Seattle, WA, US
Interview date: 2025-12-15
Got offer: False
I interviewed for a Machine Learning Engineer position. The interview process consisted of two phone screen rounds and four onsite rounds.
Phone Screen Round 1 (Coding): The question was to implement a Morse code encoder and decoder with four progressive parts:
The key idea for Parts 3 and 4 was to preprocess the vocabulary into a morse_pattern → set(words) mapping and then use backtracking recursion. I completed all four parts. The interviewer was quiet throughout but mentioned at the end that "most candidates don't finish Part 4."
Phone Screen Round 2 (Tensor Programming): I passed this round. I don't remember the specific questions, but it mainly tested GPU-related conceptual questions and simple implementations.
Onsite Round 1 (Coding - Schema Validation): I had 60 minutes to implement a schema validation system. Given a JSON-like data structure and a schema definition, I needed to verify whether the data conformed to the schema. This required recursively processing nested structures. I used dataclasses to model the schema and spent considerable time on the design. I got the core recursive logic correct, but didn't cover all the edge cases. The interviewer, a Principal SDE, was serious throughout.
Onsite Round 2 (Project Deep Dive + Behavioral - HM Round): I discussed my past projects with the Hiring Manager, focusing on my experiences with agent runtime and evaluation pipelines. The HM asked about my experience with distributed systems, which I admitted wasn't extensive – potentially a negative point. Amazon Leadership Principles-related behavioral questions were also interspersed throughout the round.
Onsite Round 3 (Paper Read - DeepSeek-V3 Technical Report): This round was interesting. I was given the paper 48 hours in advance and had a 30-35 minute technical discussion + 20-25 minute behavioral portion during the interview. I prepared a 7-point framework: Problem → Key Idea → Architecture → Why These Choices → Evidence → Limitations → Production Translation. I discussed the core innovations such as MLA (Multi-head Latent Attention), Aux-Loss-Free MoE, MTP (Multi-Token Prediction), and DualPipe. The interviewer mainly asked "why this design?" and "how to use it in a production environment?"
Onsite Round 4 (System Design - Distributed Data Processing Pipeline): I had to design a distributed pipeline for training data preprocessing. I drew a relatively complete architecture diagram (S3 → Kafka → tokenizer → deduplication → quality filter → output). However, I was not deep enough in my tradeoff justifications when questioned. The interviewer, a Principal SDE, expected to hear deep analysis such as "where it would break at scale" and "why choose Kafka instead of X."
Final Result: I was rejected. No specific feedback was given. I suspect the reasons were: the coding round wasn't strong enough, I lacked distributed systems experience, and the system design didn't demonstrate sufficient Staff-level depth.