MOLOCO ML Engineer Intern Screening Interview: It Felt Like It Went Well, and I Was Still Rejected
Loading comments...
Loading comments...
After passing the coding test I had the screening interview for MOLOCO's three-month ML engineer internship. It ran a little over an hour on video: motivation and role understanding first, then a long, deep pass through my projects and ML and CS fundamentals. I was rejected.
The interview looked at ML knowledge, CS fundamentals, and whether I fit modeling or infrastructure better.
Motivation and the role. Why MOLOCO, what I would do if rejected, how I understood the role. At one point the interviewer said my description of the role did not land at once and asked how I had read the posting and the engineering blog. I added what I had taken from them about throughput and latency in large real-time ad systems.
Hanwha UGV anomaly detection. The longest part: the goal, the constraint of almost no labels, why rules decide first and Isolation Forest joins in, and how rules and the model together cover the empty period right after start. Principle questions followed, such as why Isolation Forest isolates anomalies easily. When asked whether "built the model" meant the whole system or the model alone, I scoped it: I built a model-based pipeline for an unlabeled setting and took it to an operable level.
CS fundamentals. Processes and threads, synchronization and deadlock, which I connected to the multithreading and queues I used in the internship.
LG Aimers. How boosting works, XGBoost versus CatBoost, how SMOTE-ENN differs from randomly dropping majority samples under extreme class imbalance, and which metrics I watched.
Google ML Bootcamp. On fine-tuning with QLoRA and 4-bit quantization on a small GPU, how the base model and LoRA combine at inference.
Scenarios and tools. How I would build ad-conversion prediction hackathon-style, and how backpropagation works in PyTorch.
Strong: depth of preparation. I could explain UGV anomaly detection end to end, and boosting and imbalance handling from LG Aimers in detail rather than as one-liners.
Short: an audio problem early on made me repeat myself, and my first description of the role did not land at once.
It felt like it went well. The conversation ran long and I answered most questions without stalling. I was still rejected, and since I do not know why, I will not guess.
What I take from it: if I fail where I felt I did well, my bar and the company's are still apart. CatBoost internals, how quantization changes computation, and ad-conversion prediction are the areas from this round I will dig into past talking-level knowledge. And practicing a thirty-second answer that explains the role in one go.