Cisco · Primly Community

Cisco machine learning engineer interview: what to expect and what they actually care about

ml_mike · 4 replies

Did the Cisco MLE loop for a role in their AI/ML group that's working on network anomaly detection and predictive maintenance. Not the sexy LLM stuff but actually a pretty interesting domain if you're into applied ML.

Here's the loop structure as I experienced it in 2026:

Technical screen (45min): ML fundamentals. Precision vs. recall, when you optimize one vs. the other, class imbalance handling. They gave me a scenario: you're detecting network intrusions, false positives are annoying, false negatives are catastrophic. How do you tune your threshold? Walk through the math.

Coding round (60min): not LeetCode. More like applied ML coding. I was given a dataset description and asked to write code to train a baseline model, evaluate it, then suggest improvements. Python, sklearn, pandas. Reasonably comfortable for anyone with 3+ years of hands-on ML work.

ML system design (60min): design a real-time anomaly detection pipeline for network traffic. 10M+ events per minute, need sub-second latency for alerting. This was the round I found most engaging. We talked about feature engineering for network data, online vs. offline learning, model serving infrastructure, how to handle concept drift when network attack patterns evolve.

Behavioral (45min): STAR. A project you owned end-to-end, a time your model failed in production, how you've communicated uncertainty to non-technical stakeholders. That last one comes up a lot at Cisco because you're often presenting to network ops folks who don't speak ML.

Research/domain round (optional, they called it): since I'd published a paper in anomaly detection adjacent area they added a 30min round that was basically a technical chat. Not sure if this is standard or was specific to my profile.

What mattered most: system design and production mindset. They're not building hobby models. They want engineers who've seen models break in production and have opinions about monitoring, retraining triggers, and evaluation in a streaming context.

Comp for senior MLE in San Jose: my offer was around $195k base with the usual bonus and RSU structure layered on top. Felt fair for the level.

4 replies

ds_dmitri

the 'false negatives are catastrophic' framing for the threshold question is exactly the kind of business-context ML question that trips people up if they've only practiced the abstract 'just tune the threshold' answer. it's not enough to know what recall is, you need to be able to reason about cost asymmetry in the real application.

numbers_only

$195k base senior MLE San Jose 2026 is noted. below recent FAANG rates but above most enterprise tech companies of similar scale. makes sense for the domain.

hardware_hugo

the fact that Cisco is building real-time ML on network traffic at 10M events/min is actually impressive. embedded folks do this kind of constrained inference problem all the time but seeing it at the networking layer is interesting.

ml_mike

yeah the domain is genuinely interesting if you don't need to be at the frontier of model architecture. it's applied ML with hard real-time constraints and a lot of feature engineering creativity. not research but not boring either.