Bart Farrell: All right, first things first, who are you, what's your role, and where do you work?
Nic Vermande: Hi, my name is Nic. I'm currently with ScaleOps. I'm a senior developer advocate here, where my role is essentially to be the face of the company. So on the one side, educating people around Kubernetes resource management, so things like traditional CPU memory management, and nowadays also about how to do management of GPU and AI workload. There's a lot of confusion about that currently. And also the ecosystem is evolving quite fast. So of course, I need to educate people about that. And also how ScaleOps is just closing the gap to be able to scale all these resources and run in production at scale.
Bart Farrell: Now, which Kubernetes operational problems genuinely benefit from AI and which are still better handled by deterministic control loops and conventional automation?
Nic Vermande: Anything binary and time-critical things like bin packing a pod, acting on a taint, or enforcing a disruption budget stay deterministic controls because it's fast, it's well understood, and a probabilistic layer bolted on just adds a new failure mode for essentially no benefit. On the other side, things like pattern recognition really earns its place in judgment calls, buried in noisy signals. Things like, is this memory growth a leak or a daily cycle? Is this spike predictable? What does normal look like for this workload over a month? HPA reacting to a CPU threshold after latency already degraded is the classic case for the second point.
Bart Farrell: Now, what context does an AI system need before an operator should trust its recommendation? Telemetry, deployment history, workload behavior, SLOs, policy, previous incidents, anything else comes to mind?
Nic Vermande: Everything in the categories you listed. I think that's a good, extensive list. Everything is correct, but there's one thing people underweight all the time, which is actually time itself. So a recommendation built on a week of data or a synthetic test isn't trustworthy no matter how much telemetry has been fed into it. So you need reproducible behavior across load and concurrency variation and that workload's own history rather than a fleet average. So current SLOs and policies as hard boundaries it can't recommend past and whether a similar incident happened before so it isn't blind to a failure you've already paid for. So all of that is really important over time. And that's what makes the difference between what should be done manually versus what should be automated. So over time, it's difficult. Software is changing over time. You have version 1, version 2, multiple releases per day. If you need to change things like resource allocation based on that, for a human being, it's not really possible. So you need some sort of system that is able to understand all this utilization over time, all those changes, where they are coming from. This is what ScaleOps does because there's no real solution on the market or in the ecosystem that would do this for you. And at the same time, we have communities that have a lot of visibility, but when it comes to automation, safe automation based on those patterns is quite difficult to do.
Bart Farrell: Where should human judgment remain in the loop? And what would an AI system have to prove before you let it move from explaining or recommending a change to actually making one?
Nic Vermande: Human judgment stays on anything that is novel, like the first policy decision on the workload or a crosscutting change, live incident, because that's exactly what the system hasn't been validated against. All these things that are unpredictable. What really earns the move from recommending to acting is a track record, not a demo. We want it read-only first, compare with what it would have done to what actually happened on that workload across real variation, long enough to trust the pattern and keep automatic undo for when it's wrong. So explaining is sort of cheap, but acting without a working undo is what has to be earned. Again, here I'm going into more like the ScaleOps, how it works, because this is what is lacking in Kubernetes, especially when it comes to resource management. So it's easy to get you overviews and dashboards and observability, visibility, all of that, but being able to automate at scale in production while making sure it's both reversible and production-ready, so it doesn't mess up with critical workloads, is very difficult to do. And you need a real platform to do that, where that is very smart and intelligent, so you don't break stuff.
Bart Farrell: Nic, what's next for you? And if people want to get in touch with you, what's the best way to do that?
Nic Vermande: The next big event would be KubeCon, which is going to happen in Salt Lake City. People can find me there at the ScaleOps booth. We'll have a huge booth, so you can't really miss us. And in terms of how to reach me, I'm a runner, so Strava is still a good option. But I mean, more seriously, probably LinkedIn. You can get in touch with me. I'm also on X. And if you want to talk to me directly, you can send an email at nic at scaleops.com.
