Replit · Primly Community

Replit machine learning engineer interview: what rounds to expect, what they're actually testing

ml_mike · 4 replies

went through replit's ML engineer loop a couple months back. didn't take the offer (leveling disagreement) but the process was interesting enough to write up.

for context: i'm 8 years in ML, mostly NLP and latent space stuff. replit is doing serious ML work on code generation, code completion, and understanding developer intent. it's not just 'slap an LLM on it' territory anymore.

the process:

recruiter screen, then an ML phone screen with one of their research engineers. that's where it started to differentiate from a generic loop. they asked about: how i think about latency vs quality tradeoffs in model serving, my experience with fine-tuning or adapting foundation models, and one question about how i'd approach evaluating a code-completion model (where 'eval' is genuinely hard because there are many valid completions).

the eval question is the one to prepare for. they're not looking for a textbook answer. they want you to sit with the messiness of it. what metrics matter, what they don't capture, how you'd do human eval at scale.

onsite rounds: ML coding: implement something from scratch. in my case it was a simplified beam search decoder. numpy, clean code, explaining tradeoffs as you go. ML systems: designing an inference pipeline for code completion at low latency. we went deep on model quantization, caching KV state, batching strategies. they knew their stuff. research discussion: they gave me a recent paper (around code LLMs / retrieval augmented generation) and i had to present tradeoffs, what i'd try next. more of a technical conversation than a grilling. values / working style: pretty standard behavioral stuff.

leveling: they were thinking L4/L5 for me. the gap was about scope expectations for someone with my years. not uncommon.

comp they mentioned for L4 ML was roughly $185-200k base with equity, remote possible but they prefer bay area for ML specifically.

if you're coming from pure research (papers but no production), you'll want to show you understand inference constraints and deployment realities. they're shipping product, not writing papers.

4 replies

content_cole

the eval question for code completion is genuinely one of the harder interview problems in ML right now. there's no clean answer. did they seem to want you to land on something specific or just explore the space?

ml_mike

explore the space, clearly. i started with pass@k and they immediately asked what it misses. then we went into human preference evals, edit distance to accepted completion, latency of acceptance. they pushed back on each thing, not to trip me up but to see how i reason under pressure. good conversation honestly.

remote_swe_42

the beam search implementation in an onsite is interesting. how strict was the time constraint? did you have to get it fully working or was pseudocode with explanation acceptable?

infra_ines

the inference pipeline design round sounds like the most interesting one honestly. KV cache strategies for code completion is a real problem, not a toy question.