Transparent Trust: What If AI Showed Its Work?

So long story short, I am an auditor. So I'm actually like most of y'all's enemy. And I apologize on behalf of my profession. We have a really bad name. But, part of being in GRC in audit, I got to see a lot of applications come through our door at the company I worked at, a lot of AI tools. And I got to try to company I worked at, a lot of AI tools. And I got to try to challenge to tackle the challenge of something called like, we call AI governance. And you probably hear that thrown around, like, a ton. But, like, what does that mean to govern AI? Like, does that mean anything to anyone in the room Reddipalli, truly? But as I was, you know, working with different teams at my organization, I realized that a lot of AI, everyone's trying to make it accurate. Do you think accuracy equals trustworthiness is a question I have for everyone. No. If AI is accurate 99% of the time, but you never know when it's not, can you trust it? So, and that's what we were learning was we couldn't tell when it was not gonna be right and we couldn't tell why it wasn't right. And I'm a very stubborn and angry person generally, so when the AI wasn't doing what I wanted to do and I had a colleague tell me, I bet you could do something with Cloud Code, 2 months later, here we are. And so this app, which, it's an app, it can be deployed to AWS, it can run locally, is an idea of building a knowledge layer that AI uses, and then it defends itself. It explains itself. It tells you how it uses the knowledge to give an answer. That's how the confidence is drawn. But it can't just, you know, go wild. You're not sending it on the Internet. You're not sending it in all of your documentation. You know, I think a lot of AI today is, yeah, just hit everything up to your AI and feed everything into it, and it's going to figure it out. I think soon. I don't think today. So the idea is that you have, I call them skills. So this was birthed out of a Claude project. Claude told me to build a skill. I don't think that's what I really built, but that's where we got the name from. And the idea is you can take multiple sources of information, send them to an LLM, and it starts to build a knowledge layer and that knowledge layer becomes your intermediary, it becomes your safeguard. You can obviously prompt like let LLMs go past that, but it gives you that foundational knowledge that you otherwise wouldn't have. So URLs are the safest thing I can demo to you right now. You can obviously use documents, you can use Slack, Notion to create this knowledge base. I'll start with the URLs. The way it works is you can stage them. Let's see. And so this is all publicly facing docs that I have, that we have on our website. And what you do with those is you then go here and generate. Alright. So it came back. So it basically took the 3 URLs I gave it. It took all the information in those URLs and it turned it into this like piece of knowledge that's gonna go into the knowledge base and it's gonna then be referenced in the future. It also will help to cut down processing time. The scope at the bottom is kind of forward thinking in the idea that, eventually you won't have to walk through this process. The scope will determine future additions to the knowledge base. Okay. We'll just do that. And then it's saved. And so this is like a few I've put in so far. But what you can do with it is then when you go ask a question, tell me about regional hosting. That's gonna take another minute. It will not go to the web and it won't go directly to the source that I uploaded. If I say it was a document or anything. It'll go to that knowledge layer. So the other reason I wanted to build, and here we go. And so the other thing you can do, so right now because everything is now in my control, I'm like that's a really lengthy answer. I'm going to go back to the prompt and start to tweak it. Where Where is it? So here. So here are all the prompts that are currently like in use in the platform. So I the other thing I am big on is no black boxes with AI. There's a lot of tools where they allow you to layer prompts on their system prompts. But when you do that, now you could potentially be worrying with their system prompt. You don't know why. Right? You don't know why it's giving you a bad answer because you don't know what they're using. I have a big belief that prompts 1 day will be like Excel formulas, but not today. So, so yeah, so let's see. Go back to RFPs. So the other thing about the curated knowledge layer is something I was realizing, obviously, at organizations, you have a lot of content. You have a lot of documentation. Don't know about y'all, but sometimes it's not always the same or people have different answers or, you know, you start to war with each other, even unintentionally, arguably. So the idea about the knowledge layer is that once it's there, that's the source of truth. You're not having duplicate pieces of knowledge. And it can also tell you when you have duplicate pieces of knowledge. And the other thing is it can also take like Zendesk tickets, Slack messages, and transform that into either a knowledge layer and then beyond that documentation. Everything flows up and down. And yeah, that's the demo I have without breaking any NDAs at my current job because I can't show much else off. But the idea is also that you can use it to power other things like a chat, like a collateral building. You could then take the knowledge layers that you've built and actually just use them directly in cloud. You know, right now, I think it's like, if I'm using, data terms that I don't understand really well, the semantic layer, context layer, right? You know, there's a lot of companies dumping large amounts of data into data lakes and data warehouses and using that to train LLMs. I think the future, at least in terms of this, isn't training LLMs, but giving it the correct information. You know, my colleague and I have been joking a lot that you got to give it vegetables. We keep we keep throwing clod into a candy store and wondering why it's giving us bad responses. So you got to you got to teach it how to read. You got to teach it how to, you know, give it the right things to grow. It's it's far too humanlike to treat it like it's not humanlike in a way. And so, yeah, I was going to also demo like you can the idea is like you can take anything. So like I could have started to build a knowledge base off the dot, off the URLs from voxel 51. Like it's very transferable. It's it's this idea of building that knowledge layer. And from that you gain the understanding. And, and then wait, I realized I didn't show the 1 thing that was really important. When the AI gives its answer, let's see. Will it show up? There we go. It has to give the transparency. Oh, that broke. All right. Let's see if the other ones gave it. Let's see. Nope, it is dead, Okay? Okay, but when it's not broken, it gives its reasoning. It gives it sources. It explains itself. Yeah, and that's it.