ServiceNow · Primly Community

ServiceNow machine learning engineer interview: what the loop looks like in 2026

ml_mike · 5 replies

I went through the ServiceNow machine learning engineer loop in February 2026, for a role on the Now Assist team (that's their AI-powered workflow automation product). Sharing because the ML interview landscape is fragmented and company-specific variation matters a lot.

First: what they're building

ServiceNow's ML team is doing applied AI, not research. The focus is on LLM-based workflow automation, classification models for IT service management (incident categorization, change risk assessment), and retrieval-augmented generation for their virtual agent. If you're coming from pure research or academic ML, this is applied engineering with a heavy inference-and-serving component. Know that going in.

The Loop (5 rounds)

Recruiter screen, ML coding screen, ML system design, applied ML case, and a behavioral/leadership round.

ML coding screen

60 min, two parts. Part one: implement a simple attention mechanism from scratch in PyTorch (not calling nn.MultiheadAttention, actually writing the math). They're checking if you understand what you're using. Part two: a statistical question about evaluating a classifier on imbalanced data. Precision/recall/F1, why accuracy is a bad metric on 98/2 splits, ROC vs PR curves. Know this cold.

ML system design

Design a system that classifies incoming IT support tickets into categories (incident type, priority, routing) at scale. I talked through: feature engineering on the ticket text, fine-tuning a smaller BERT-class model vs. prompting a larger LLM, latency constraints (they said SLA is 500ms), online vs batch inference, model monitoring for drift, retraining triggers.

The latency constraint is real at ServiceNow. Their platform runs on enterprise SLAs. A 10-second LLM call is not acceptable in many contexts. They pushed me on when I'd use a smaller fine-tuned model vs an LLM API call and I had to defend the tradeoff.

Applied ML case

Given a dataset description (they didn't give actual data, just schema and stats), walk through how you'd build and validate a model to predict whether a change request will cause an incident. This is real ServiceNow domain stuff. Think through: label quality, class imbalance, feature leakage if you include post-change features, how you'd set thresholds for a business that has high cost of false negatives.

What I'd prep

Transformer internals (you will be asked to explain attention). LLM serving patterns: batching, quantization, caching. Imbalanced classification. The Now Assist product, read about it before you go in.

Offer: Senior MLE, $230k base, $260k total cash. Remote eligible. RSUs separate.

5 replies

qa_quinn

the 'implement attention from scratch' question is a real divider. half the people calling themselves ML engineers can't actually do it. what was the expectation on efficiency, or just correctness?

ml_mike

just correctness first, then they asked about complexity. if you get the scaled dot-product right and can explain why you scale by sqrt(d_k) you're fine. they weren't expecting flash attention from scratch.

hardware_hugo

the latency constraint angle is real for enterprise. 'we just call gpt-4' is not an architecture at p99 enterprise SLAs. good they're pressing on this in interviews.

quietquit_quincy

230k base for senior MLE at servicenow seems right for 2026. big tech comparison: google/meta are higher but they also have a different interview bar and TC ceiling. servicenow is a fair trade for anyone who doesn't want the big tech grind.

returner_ren

the domain-specific knowledge requirement (knowing what Now Assist is, understanding ITSM use cases) is something I underestimated when I prepped for enterprise ML interviews. thanks for calling that out explicitly.