MongoDB · Primly Community

MongoDB machine learning engineer interview: not what I expected

ml_mike · 4 replies

went through the MongoDB MLE loop earlier this year for a role on their AI and ML platform team. they're doing real ML work now, mainly around vector search integration in Atlas, natural language query generation, and internal ML tooling. this isn't a database company doing ML theater.

here's what the loop looked like:

recruiter screen: quick, covered background and why MLE at a database company. valid question honestly. I said something about how the intersection of ML and structured/unstructured data retrieval is where I wanted to be. they seemed to find that credible.

ML technical screen (60 min): I got a question about designing a recommendation system with a focus on the retrieval stage. relevant given that MongoDB is pushing Atlas Vector Search hard. we talked embeddings, approximate nearest neighbor search (they know HNSW since it's in their product), evaluation metrics. solid conversation.

onsite had 4 rounds: ML system design: design an entity resolution pipeline. we talked feature engineering, candidate generation, blocking strategies, model choices, and how you'd handle a cold start. they wanted end-to-end thinking, not just the model. coding: LC-style, one medium. not ML-specific. just a standard algorithm question. you need to be able to code cleanly, not just think in PyTorch. applied ML / case study: given a business problem (something about improving query suggestions), how would you approach it. very product-flavored, not a pure research question. they care if you can scope, not just model. behavioral: ownership, cross-functional work, and a time you shipped an ML system that didn't work as expected and what you did. the failure question always matters more than you think.

timeline was about 5 weeks total. offer was competitive for a senior MLE: NYC base around $220k, significant RSU package.

if you're coming from pure research, you'll need to show you can ship. if you're coming from industry ML in adtech or fintech, the vector search and retrieval angle is the most natural on-ramp to talk to.

4 replies

ds_dmitri

the entity resolution pipeline question is a good one. often separates people who know the modeling part from people who've actually built production pipelines. blocking alone is a whole conversation.

infra_ines

curious if they asked anything about how the ML models get deployed or if it was purely the modeling side. do they have a platform team that handles infra separately?

ml_mike

they mentioned a platform team handles a lot of the serving infra, but they still expected MLE candidates to understand deployment basics, latency constraints, model versioning. not expected to build Kubernetes yamls from scratch but not totally shielded from it either.

remote_swe_42

the vector search angle makes sense. Atlas Vector Search is their main AI push right now. if you can walk in knowing that space well you're differentiating from candidates who just know MongoDB as a JSON database.