Dylan Ratcliffe

And we thought that would be incredibly helpful.

And while it's information that people didn't have before, it's also work they didn't have to do before.

And so people would look at that and go, are you telling me I have to review these 500 different things to see how my change is going to affect them?

That is that's terrible.

That doesn't help me at all.

And so that was a pretty big realization that that the whole thing that we had built, the whole real time discovery engine

was not helpful.

In fact, actively unhelpful to show to a human.

It was, we were actually in Greece at the time.

The team is fully remote, but we get together every now and again.

And like a few times a year, and we're all in Greece at the time.

And GPT-3 had just, we just got API access.

We'd been on the wait list for quite a long time.

And we were doing a bunch of experimentation with what you could do with the data set that we already had, the features that we already had, that we realized that

nobody wanted because it was actively unhelpful.

And what can an LLM do with that?

And it turns out that having a directed graph that it can traverse in real time of all of the dependencies and having certainty of it is the perfect thing for backing an LLM.

It means that it's not possible for it to hallucinate because we're telling it what the dependencies are, but it can still reason about impact and risks and things like that.

And so

When we had that realization, our roadmap was suddenly we are going to build a product that tells you what the risks of your deployments are.