Berk Yilmaz

So the AI changes the speed.

We know that.

But the verification and the audit trial change whether the results are actually usable for dysregulated customers.

And that's the equation actually we're solving at Sentinel.

Yeah, this is something I think about a lot, and it directly shapes how we architect the Sentinel.

The dominant approach was, in AI tooling right now, the biggest available model sends everything all the time.

And that works fine if you're in the cloud, and latency is nice to have.

It completely falls apart when your customer is operating an SICP or an air gap facility where the model runs on a local hardware and every watt matters now.

So our approach is you don't use AI where we don't need to.

Our compliance checking, for example, is entirely deterministic and no large language model is involved.

We map the findings to control frameworks using a rule engine.

not a language model, because a regulator doesn't want to hear your compliance assessments was probably right, right?

It needs to be the same answer every time.

So for this, we have to get a deterministic approach.

On the model side, you're designing for right sizing, using smaller, faster models for orchestration and routing, and larger model only when the task actually demands it.

The goal is that the system gets more efficient as it matures and not less.

The more you use it, more of the pipeline hits the cache, the fewer expensive models calls you need.

And efficiency isn't a performance optimization for us.

It's a deployment requirement.

If it doesn't run on custom hardware behind the locked door, it doesn't ship, basically.