Hello, everyone. Welcome to Tales from the Trenches Real AI at InterWorks. My name is Derek Austin, and we're gonna dive in to some Real AI stories that we've worked with, things that we've done here at InterWorks engagements and stuff we've done internally as well. But before we do, I want to go back in time just a little bit, back in time to the year nineteen seventy five. In nineteen seventy five, a twenty four year old Kodak engineer named Steve Sawson built something that no one had ever seen: a digital camera. It was the size of a toaster. It weighed about eight pounds. It had a whopping zero point zero one megapixels, and it took twenty three seconds to save a single black and white photo, but it was a digital camera. It recorded to a cassette tape and it was able to be reproduced and brought like modern digital cameras. He brought this invention to Kodak's leadership. They looked at it. They knew what it was. They knew the importance of this technology. But they did something surprising. They more or less told him to bury it. They were terrified, not because it didn't work, but because it worked too well. If digital photography took off, it meant that it would cannibalize a lot of Kodak's film sales. And film was huge in the Kodak empire. It equated to billions of dollars a year. It was a large part of their business. And so for thirty years, they sat on this invention. Kodak watched as people like Canon, Sony, Nikon, and eventually every smartphone manufacturer built the future that could have been theirs. Something that they had invented and that they could have utilized, but they just sat on it. In twenty twelve, Kodak filed for bankruptcy, as you might know. Today, this prototype, as you can see in the image here, it actually sits in the Smithsonian. Kodak didn't die because they lacked innovation. They died because they saw the innovation and they sat on it. They didn't want to take advantage of that. They saw the future and they chose not to adapt. And so as we look at AI today, we have a question as well. And the question isn't can we innovate? It's, will we act on what we already know? Every so often there's something that comes out that changes everything. If you look back in history, we have things like the printing press, the steam engine, electricity, the Internet. AI, depending on who you ask, is shaping up to be that big. It's going to be huge. It's not just a new tool. It's an entirely new category of tool. It's huge. It's revolutionary. It's our moment of electricity if you will. Will it be that big? Who knows? But it is shaping up to be very very large and very monumental for our generation. As we dive in, my name is Derek Austin. As I said, I'm the Curator Platform Lead here at EnerWorks. I basically have been in EnerWorks for thirteen years and spent the majority of that time in software development, building an internal product that we call Curator. And during that time, I've seen a lot of different things come and go, a lot of different technologies. I would say one of the things that's most exciting is the AI revolution that we're in, in the midst of right now. There's so much that's happening around us and changing, even from a couple of weeks ago, a couple of months ago, but especially if you look back a year or two. The world is evolving and changing rapidly, and we work to go with that flow, to be in front of it, to know what's changing and how we can take advantage of it in our day to day lives. Interworks has always had a mantra. If you don't know much about us, our mantra is doing the best work for the best clients with the best people. Through that, we've done a lot of different things over the years. We started as an IT services firm in Little Stillwater, Oklahoma. But over time, we expanded. We did more and more in that space, but we moved on to things like data analysis as well and database work with partners like Snowflake. I've been a part of Interworks for thirteen years, like I said, and I've seen a lot of this rise and fall in our company. And we are truly a global brand at this point, doing a lot of fun stuff in a lot of different areas, everything from your mom and pop server installations all the way towards working with global brands across the world. Outside of InnerWorks though, as you guys are aware, there's a lot happening, a lot happening in the tech space. If you go back to November of twenty twenty two, you guys probably remember a monumental shift that happened almost overnight. OpenAI released ChatGPT on November thirtieth, twenty twenty two. It was this free chat built free chatbot built on GPT three point five at the time. Within five days, they had a million users. Within two months, one hundred million users. For just a minor comparison, it took TikTok nine months, Instagram two point five years, and Netflix ten years to hit those numbers. It wasn't just another tech product launch. It was a cultural event. If you look at ChatGPT, if you look at chat based AI these days, it's not just the tech people that are using it. Your parents are using it. Your kids are using it. Your CEO is using it. Companies scrambled. Some banned it. Some ignored it. Some leaned in. But as we fast forward to today, the question is no longer will AI matter, it's how fast is this moving and what does it mean to me? As we moved into twenty twenty three, we realized this is real. This is something that will drastically change all of our lives. This is when you see things like GPT-four coming out and passing the bar exam, scored in the ninetieth percentile. It passed AP exams. It was writing code in dozens of languages. And we see this shift from how useful will this be to what can't it do. Companies started integrating AI in literally everything. Microsoft put Copilot in their Office suite. Google launched Vard. Anthropic released to Quad. Organizations are split. Still, half are banning it, half are requiring it. And the honest ones are admitting they're not really sure what to do, but they know that something big is happening. They're realizing that this isn't autocomplete. It's not another Google search. It's something that reasons. It's something that creates. It makes connections. And it's getting better every few months. Moving forward just a little bit more to twenty twenty four. It is a big shift, agentic shift, if you will. It goes from AI being a chat assistant to really being an autonomous worker. Before you ask a question, AI gives an answer. You just kind of do the work and piece it together. You describe a goal, AI breaks it into steps, executes them, delivers results. Twenty twenty four really was the year of the AI agents looking back systems that can browse the web, write and run code on its own, call APIs, read documents, chain together multiple steps, all with you just kind of babysitting and sitting there in the background. We saw various tools pop up that took advantage of this, things like AutoGPT that just really kind of stepped in and leaned into that. Quad with tool use would be another example of that. Custom agent pipelines started popping up and showing up in real production workflows. The mental model really switched at this point. AI wasn't just a search engine. It wasn't something that we talked to. It became basically a junior employee that we could delegate to, that we could work with, that we could reason with. And, of course, the last year and a half, two years, that's really just been accelerated. We've moved into more of a software factory model with a lot of companies, a lot of things that are happening out there. A lot of people are diving into things like quad code, cursor, and devin. They're not just assistants, they're builders. QuadCode sits in your terminal. It reads your entire codebase. It writes features, runs tests, fixes bugs, opens pull requests. You simply review while it ships code. Startups are launching with two or three people and doing the work that used to take twenty, not because the work disappeared, but because AI accelerates that much. It helps with that execution layer. The pattern is still the same, though. Humans are setting the direction. AI builds. Humans review it. AI iterates over and over and over again. The ratio of thinking to typing really has completely inverted at this point. And this isn't theoretical. For a lot of people they look at this place and they say, Wow, that's really cool. They see other companies doing that, but really it's something that's approachable for most companies with a little help. If you look back on my career, if you pulled up my LinkedIn page, one thing you might notice is I've been programming for a very long time. I've written code more or less for twenty five years, and I've been across a lot of different languages, seen a lot of different technology come and go and disappear. But over the last couple of months, maybe even year, it's been a long time since I manually touched a line of code. My programming job isn't gone. It's evolved. It's changed. It's faster. I'm making more things and having more output than I ever have. I'm having a blast. But what I used to do just a couple of years ago, it's gone. It's an entirely new landscape at this point. Completely different. Technology, programming, it's just the first on a curve of a lot of different things. Every career is going to evolve in some sort of way. The question isn't if, it's where, where you are on that adoption curve. And so I want to dive in today and talk about that curve and where you find yourself on it, but also the next steps and how you can kind of move forward in your AI adoption journey. If you think about AI at a very base level a few years back, the biggest thing was actually using it. I feel like a lot of people are, if you're on this webinar, you probably are already using AI. But for a lot of people that was a big barrier for a long time. Most people would hear the words AI and they'd overcomplicate it. They wouldn't even try because they assumed it would be hard to use. The thing is, we already have access to AI. There's literally free tools out there. Most people have some kind of enterprise thing at their various companies as well, but it's very, very easy. The ease of access is just easier than it's ever been. You can use AI as quickly as you can sign up for a new email account. To get started, all you have to do is literally open a chat window and describe your problem, and it figures it out. It takes it from there. That's it. There's no specialized accounts to configure. There's no custom setup, no code. You just hop in and have a conversation. And so what does this look like? A few years ago, lot of people were really struggling with this piece. And I think the thing is people really try to overcomplicate it. And so I have this kind of funny personal story that I like to tell. This was in the early days of AI. My dryer died one day. And I started to call the repair guy, but before I did, I got to thinking, I wonder, I wonder if AI can take care of this problem. The dryer was running. It just wasn't heating anymore. It was simple to close all day and they'd just be a wet, soppy, cold mess without any heat. And so I opened up the ChatGPT app on my phone. I started describing the problem. It asked me a few questions. I answered. We went back and forth. And we didn't really get anywhere at first, But it had one line that was really special to me. It said, If you're okay opening up the dryer, I have some next steps. I'm not a dryer repair guy. I'm technical, but I don't spend my days studying electrical engineering books. I didn't know what I was doing in there more or less, but I was up for a challenge. The dryer was already broken. It wasn't going to get worse. And so I followed its instructions. I took out my multimeter. We started taking things apart. I basically took pictures. I sent it to the app. It told me exactly what to test, how to test it even, and what to replace. And in the end, I ended up fixing my dryer. It's still running today. No experience, no user manual, just a conversation with AI and a thirty dollars part, which honestly, I got the link directly from ChatGPT for that part. I didn't even have to go find it. It linked me directly to it. This story, it's kind of a silly illustration, but here's the thing. Everyone has something that's this approachable with AI. I'm not an appliance repair guy. I'm a software person. This is something that's entirely out of my depth. But with a tier one AI solution, I hopped in with no setup, no custom instructions, no code. I just had a conversation. AI was patient, specific, and walked me through exactly what I needed to do in a situation where I'd never found myself, kind of like a good tutor. It met me at my level and it walked up, taking it one step at a time. But the thing is, I needed to get there. I needed to ask. A lot of times I feel like even as people become more familiar with AI, they're asking the question, should I use AI for this? Or maybe even, is it possible to use AI for this? And a lot of times that's where the conversation stops. They don't realize that AI can do things that are maybe a little bit outside of its box. Maybe they'll ask it for computer tech questions, but that's the end of it, the end of the line. In each of our jobs, I feel like there's a lot of things that can be automated with AI that we don't realize are there, things that don't, on the surface level, seem like something that AI could approach or AI could do. If you think about the dryer illustration, AI doesn't have hands. It doesn't have any sort of way to interact with that dryer. But it had me, and it could tell me what to do and how to test things. And then it did the debugging that I couldn't. And so I would challenge you guys as you approach AI to not ask, should I use AI for this? But more of, can I? Sometimes it works really, really well. And honestly, sometimes it won't. But through that process, learn what works and what doesn't. And you'll find things like my dryer repair that you never expected AI to be able to do that's actually really, really good at. And so looking at that in a business sense, I have an interesting example here. And I'll have a lot of examples as we go through today, but this is one that really stood out to me and I think a lot of people can resonate with. I was working with a client and they are a client that acquires clients of their own or companies of their own or a company that acquires companies, and they had inherited this crazy spreadsheet. It had color coding and multiple rows and the client data, the customer data, if you will, was a big four block chunk, and things were hard coded in there in really weird ways. And as we looked at this with Python trying to get this into their system, their actual customer database, we ran into some interesting issues. It was just going to be really tough to process this with straight line code. And so what we did is we kind of described this spreadsheet to quad code, and we told it what was going on and how it was laid out, different business rules that were built in there. And before long, AI built us this Python parser that would go through and convert this spreadsheet into a very simple line by line, one row per customer file that was able to be imported into any system. If you think about this, not everybody has a crazy Excel spreadsheet like this, but the task and the output are very similar, right? You're the domain expert. You have something crazy that's legacy that has been around forever and nobody knows how to approach turning it into modern things, into a modern system. AI more or less wrote this translation layer for us. We explained it. We explained how it was set up, all the business logic, and then it took it and ran. Very quick turnaround. Something that was previously, even four or five years ago, impossible or really, really difficult became approachable in just a few hours. Another example of this is something we do here at InnerWorks actually. When we have a client engagement, often we want to record that, either for internal or external use. And recording information about engagements before, before AI, was something that happened only so often because it required a lot of manual work, somebody to document what happened, the different approaches taken, the clients that we interacted with, the actual people, the names, that type of stuff. It was a lot, a lot of different pieces that needed to come together, that needed to be processed, that needed to go into making this document that was kind of a handoff document to other developers coming behind us, or maybe even a case study that's that's web focused, which just took weeks, a ton of time. I think marketing told me six to eight weeks. So long. But with AI, we're able to take a simple Zoom call transcript and turn that into a case study with some pre processed instructions. We have a draft in just a few minutes. We take that one transcript from that call and we process it. We pull in different information like any text documents we have about that. And we can create things like a one page document on what happened or a LinkedIn post with a case study about what we did with XYZ company, maybe an email snippet, some slides with bullets, just some information that's very easy to tell people what happened and how it worked. Of course, AI is never perfect, and so we have the system flag things that it doesn't know for sure, things like this piece needs verified or it needs clarification, and of course marks the draft if it doesn't email. But the thing is the bottleneck was never collecting the information. It was processing it. It was that conversion, if you will, to the next step. And so as we look at that, that moves us really into tier two, the next step of AI. Tier one, it's very easy. It's chatting. It's kind of having that back and forth. I think most of us, not all of us, have interacted with that. But the next step of AI is where it really starts getting interesting. That's where you're creating things where we give different processes job descriptions, if you will. This is where AI really started getting interesting to me. We're creating standing instructions once, and now anyone on the team can run the same flow and get a consistent high quality output without knowing anything about the prompting or the pieces underneath, basically creating reusable flows, if you will. A really interesting example of this for Curator right away was release notes. As we release a new version, we could take all of the information that developers had put in about a specific change, and we could ship that out through a processing system to create a client friendly release note page. So how do you do this? There's a lot of different ways to create these custom kind of AI assistants depending on what you're using or how. QuadCode has things like skills, for instance, or commands. In GPT there's different ways to approach this as well. One of the first ones that came out was what OpenAI called custom GPTs. And it was a very friendly user interface where you could simply define things like a persona, a name, rules, things that this AI should do or not do, and these repeatable flows. Essentially, it was just like chatting with a chatbot, just you're giving the instructions to this GPT, and it will remember it and create this processing widget. An example of this in Enerworks outside of the release notes is company announcements. We want to have a very similar brand voice internally whenever we do announcements, whether that's a middle manager or the CEO. And so we have a system like this that will process your kind of thoughts around an announcement or maybe your draft and really update it with some brand guidelines. But as you look at this, the unlock becomes really clear. You don't need a super technical user or even a semi technical user to use something like that. At the end of the day, anybody can go in and they can dump some text in this custom widget. And on the outside, the output, they get something great. They don't have to think about prompting or the different steps. It just happens. It's kind of like writing a job description. It's not a one time request. It's a standard operating procedure that allows anybody to approach AI with a very specific output. And so we've done a couple of different things that Enterworks with us outside of those, a little bit more complex. One of my favorite is this translator system experts. We utilize something called Kubernetes for some of our curator back end, which is, if you're not aware, a very complex server infrastructure setup. Uses containers. It does a lot of fancy stuff that your typical developers don't necessarily understand or know how to approach. And we have some people that are very good at that system. They're domain experts. But bridging the gap between the two was increasingly a challenge. And so we created this translator, if you will, that has all the information about the Kubernetes setup, about what's going on there, how it works, the documentation, and it can translate this information into something that our developers can work with. I think we all have this situation where no matter what the technology is there's like a domain expert and he maybe speaks in a language that the rest of us don't understand, technical jargon for a particular area. And if you think about talking to that person, sometimes you get a good answer. Sometimes you get something that's still a little vague and you don't want to be that guy. And so you go Google half of it, you give up, maybe you DM the guy again in an hour or two. But at the end of the day, you're really fragmenting that domain expert's day. All he's doing is answering questions. And so a solution for us was building a custom AI script, a custom AI GPT, if you will. We use a quad code skill actually. But it's preloaded with jargon, with technical information about that system, a glossary, infrastructure documents, the app internals, all of that kind of head knowledge that would have been that poor domain experts before is now in a space. And if we just approach that documentation with our standard like frontline developers, they would still struggle, but it gives enough information for AI to realize what's going on. And so it can take that, it can translate it from domain expert speak, and get it into plain English that a developer can work with, can understand. And basically they're able to go in, they're able to ask their systems or ask their questions and get good answers. The AI didn't replace that domain expert. He's still really important. But instead it kind of distributes his expertise so that he can focus on what he needs to instead of just answering our questions. Another really great example of this is tracking meetings. And so if you get a transcript from something like a Zoom call or a meeting with a client, a lot of times that's a lot of information. Even if somebody's taking notes, you miss things. And so you come up with this problem of, Oh, what was that timeline? Who owned that action item? Was that last week's meeting or the one before? Meetings happen all the time, more and more. And getting that information into a system that's parsable is really tough. And so something that you could do is take those transcripts, you could auto log them, you could have AI index them every conversation, every decision, every action item. In fact, Zoom has some of the stuff built in. But if you take that from there and add it into kind of a knowledge base, you can easily ask things in plain English across all your meetings. Things like, What do we say about our targets? And get instant information with a data context, maybe next steps. You don't have to have somebody say, Hey, can somebody send the meeting notes? You can easily have that right in front of you. You don't have to have that documented anywhere. You just let the system handle its things. Then decisions are traceable. You don't lose institutional knowledge when people leave as much at least. New team members can hop in. They can see the context right away. And it's that same pattern of having a persona, a set of rules, and a knowledge base, and then just something to process it. In fact, if we go back to the custom GPT agent, that first thing that OpenAI released, you can see there's actually an area to upload knowledge files. And that's kind of a crude way of, like, indexing this information, but it would certainly work. You could just upload those transcripts there. Different technology like quad code makes this a lot more accessible. You could literally just jump the transcripts into a folder. But that's something that you can do easily in that kind of tier two space. Another really interesting use of this is for triaging support tickets. And we have, I would say, have a tier two version of this, and then there's a much more advanced one that we'll talk about later. But if you think about a support flow, when a support ticket comes in, it needs some kind of basic information every single time. What is this asking? Have we seen this before? How big of an ask is this? And so what we've done with Curator is we have an AI assistant that will read the ticket, figure out what's going on, restate the ask, search past tickets for a precedent and what's going on, size the effort, and propose a solution. It doesn't automatically respond to clients. We want a human to do that. But it drops like this structured triage note. And it does it in a few minutes and something that would have taken a non call human a lot of time to work from. They can easily just hop in and see that and hit the ground running, if you will. AI doesn't replace the human. It does the homework. The human starts closer to the finish line than they would have before. It's the same tier two pattern with the persona, the rules, and a knowledge base, And it's taking it and helping enable our support agents to do things quicker, faster, and better. That moves us to this tier three layer, though. As you're thinking about AI, there's that tier one, like just getting started, the tier two, like we're building some custom flows. But tier three is where things really take off. We're not asking one AI to do everything. We're breaking it into specialized roles, specialized personas, things that hand off to each other, kind of the same way that a professional team works. One of my favorite illustrations for this is what I've deemed a newsroom model. If you think about a standard newsroom from the 80s, there's not just one person doing everything. And there's a reason for that. There's specializations. There's people that are really good at taking photos and there are people that are better at writing things down. That's just how it is. And the reason is they're very specialized in those various areas. They've been trained in those areas. And AI is really very similar. You can create flows that do various pieces, if you will. I've implemented this in a variety of ways, but a blog is a really easy way to illustrate this. Basically, a blog flow that has a pipeline of research and publishing, marketing, different steps, and even one that reviews and kicks it back to the beginning if it needs some extra work. So you could think of having a story analyst, something that looks into an idea, a research that finds facts, writes a draft, a copy editor that adds some polish, maybe an author, and then this critic role. And it kind of goes in a circle over and over and over again. And there's different ways to approach this. Different products have different ways to dive in here. One of my favorites is Claude Code just because that's what we use a lot of Claude here. But we have basically different agents within the system and then a skill that will walk through each of those and have different minds behind it with different described job tasks. Like, our critic has a very specific set of things that it's running a potential blog through. It checks all the boxes. It gives it a score. And if the score is high enough, it passes it along. If it's not, it goes back for more work in the drafting area. But the point is each agent has one job. No one agent does everything. The handoffs are very sequential and it works more or less like an assembly line. You're the managing editor. You set the idea. You set the brief. You review the output and you make the call on whether or not you use something. But at the end of the day, AI is doing a lot of the steps in between. This is the same pattern that we've seen in newsrooms for one hundred and fifty years. We're just applying it to AI. Different tasks, different pieces. And I actually do have a blog on this. We can drop in the meeting notes afterwards. But if you are interested, a lot more information on this is available on the Toric's blog under my name. And so how does this work? We can take it from just simply being a blog to being something like that Zoom transcript model that we were talking about earlier. It's the same Zoom call. We take that transcript from the client interactions, from the discussion about what happened in a project, the project breakdown, if you will, and we can extract that transcript. We can look up more information with an agent. We can research things. We can have one that drafts, one that builds out our branding and voice guidelines, and then one that does like the publishing steps as well. Moving on though, another really interesting example of this is a simple code reviewer. Nowadays everybody is using AI to build out so much code and so many new websites, people that weren't programming before. And so code review is a large part of what my team does these days because implementing these systems is really important for maintaining that big amount of code that's coming into the world. We have systems that do this automatically. We still have human people that look at things as well. But really, AI can accelerate your code review so much faster. This is one of the first things that Quad in particular, as well as OpenAI, integrated. And a lot of people have approached AI code review, but they haven't really dove in. There's a lot of different things that you can tweak and change to make it actually worth its while. A lot of people just turn it on and then they get frustrated. That doesn't give good results. But there's a lot of different things that you can do to actually make this work. In particular though, you can have it come in, you can look at the code, and you can apply your standards. And you can look at every single line that changes and have it go back and analyze that against the code standards that you want, the security standards that you want, that type of stuff very easily. In our case, we have one that does that. And then also one that looks at documentation and flags things that are drifting from the documentation. If somebody makes a code change and it doesn't line up with our public facing documentation, it complains and it says, hey, this looks like it changed this documentation area. You should also be changing the documentation along with that. And that's a really good example of how a code review agent can have a persona as well. It doesn't just have to be the standard out of the box quad code review. It can very easily be an agent that has a specific task assigned to it, a division of labor, if you will. One that worries about styles and conventions, one that worries about documentation. And then of course there's always areas that you want humans to review as well, the substance of the change and whether it's necessary. And in our work we've seen that this is really important and helpful because as more and more codes coming in, humans can't keep track of that in the same way that they used to just a couple of years ago. They need some assistance with that. We used to have pull requests come in, changes to the code, and we'd argue about styles and conventions and how people should be doing white spacing. All that type of stuff can be automated now, and we can have AI look at that and then humans just look at the important pieces like, does this make sense? Is there kind of any issue with implementing this from an infrastructure perspective? Of course, you can't just have AI in its own session review its own work. It rationalizes just like we do. And so even though a lot of the code is generated with AI now, this extra step of having an agent is an extra review that's kind of like having a second person look at it as well. It takes it out of the context that it was in, gives it a fresh context so it's looking at it with fresh pair of eyes. And even the code review agent, even if it's using the same model underneath the hood as the person creating that code, it still catches new things, which is exciting and interesting to watch. This is an example, though, of that newsroom model proven in production. Over and over and over again you can see the separation there and how helpful that is. So if we move along, the next logical step, you're outside of just building flows and you're moving on to AI writing code. A lot of you on the call probably aren't developers, but even if you aren't, this is where it gets really interesting. Because honestly, in a lot of ways, you don't have to be anymore. Sure, there's still the need for somebody that understands what's going on and the architecture pieces, the infrastructure, all of that, and it's very helpful for getting things up and running. But there's a lot of things that you can do that are safe, but also really, really helpful in the programmatic sense that you wouldn't have been able to do a few years ago if you weren't a developer. And if you are a developer, you can just light all of that on fire and just go so much faster than you've ever been able to and approach problems in new and exciting ways that you couldn't have even dreamed of years ago. Just last week, actually, there was an article published about Anthropic and how they have been using Quad behind the scenes to replace whole libraries within their systems and basically rewrite from one language to another. Massive projects that would have taken years before that can be done in a couple of days very quickly. But there's so much power there now that wasn't there before. So a couple of examples of that. One of my favorites is a conference intelligence agent. A coworker of mine showed this to me. It was one of the first things that he did with AI and I thought it was just mind boggling to even think about since I've used this model a few times as well myself. But if you think about all the different powers that AI has, now you can search the internet, you can pull together research, you can build with code. All of those things combined, you can build something that's pretty cool to look up information and kind of give you public briefings, if you will. And this is one example of that. Essentially I was going to a conference where I was speaking. I was giving presentation very similar to this actually. And as an example, I wanted to research the potential attendees, their background education, their career paths, interests, personal background, even volunteer stuff, and then rank them on how strong of a connection they'd be for me so that when we have those mixture type events, I would know a little bit about them and talking points, things that we could connect on. I had AI build out an interactive chart that I could filter, search, open on my phone, and it found a lot of things that humans would miss. Same small town connections, small town colleges, work crossover pieces, volunteer roles, and things like LinkedIn missions that align perfectly. If you think about this, though, you could swap a conference for sales prospects, partner list, a board of directors. The same workflow could be infinitely valuable, a little creepy at times, but infinitely valuable in so many situations. And so how did this look? I pulled together this is just a kind of a these aren't real people. This is a little HTML file that I had quad build with all these people in there ranked. It's a little hard to see here, but even just laying out out things like the best match, each of these attendees. And I just had Quad run overnight doing this research. I just turned it on, gave it a a goal, and let it go. And it found a lot of information, public data, various work profiles, things from LinkedIn. It's not a contact list. It's more like a strategic like intelligence brief. If you think back forty years, right, they were the old guys with Rolodexes and they'd keep track of all kinds of information. That's this but on steroids. It gives us a lot of information walking into a room of a bunch of people we've never met, knowing exactly who to talk to and why, what's important to them, what I might find important in engaging them as well. I could dive into various profiles and I could easily see connection points, talking points, conversation starters, overlaps on volunteer roles, full background briefs, specific connection opportunities, people that maybe they know that I know. And there's a lot of information that honestly, as a single human, I would never have figured out, maybe even knowing this person for years, that surfaced very quickly and easily. Moving on from that though, there's a lot of different ways that we can implement the coding side of these AI softwares. One of my favorites, though, is I had a customer who had a system that it didn't have any ability to connect to it with an API, at least not one that's logical. And there wasn't a big bulk import area. So we had to figure out how to get data in there for them. Hundreds, thousands actually, probably tens of thousands at this point, customer records that needed to be entered into the system. And the problem was, without an API, you're looking at doing that manually by hand, hiring a bunch of interns for a couple of weeks or something and hoping they don't make any issues clicking through forms. And that old answer, it might have worked. There probably would have been some bugs, some typos. But with AI, we're able to approach this type of problem in a way that honestly we couldn't have even dreamed of before. Sure, there's software that's built around this kind of robotic automation processing type flows, but programming that out takes a lot of time. What we can do, though, is we can get Quad to write that for us. We can get it to create a script that opens a real browser that logs in, that fills out forms exactly like a human would, click by click, by simply giving it some information about the page. Then we can have it run these scripts in overnight sessions. Every record's entered, validated, logged. There's no more typos. And at the end of the day, if a person can do it in a browser, we can have AI write something to do this for us. No special API access is needed or required. Of course, that's the best approach if we could get that API access, but a lot of times you simply can't. Another really interesting example if we think about coding is just approaching how we develop software in general. And this could be anything. I know a lot of people on the call might be data analysts. They might be looking at building out dashboards versus building out websites. Excuse me. But the same thing kind of approaches, the same approach works here. So I've been doing this mock up driven development. It's a term that we use with Curator a lot with Maya developers. But basically in the past, we would have had a designer look at some kind of change that we wanted to make, build some mock ups in PDF form or whatever and just like hand those out. We'd wait weeks for feedback. And then finally we'd get a developer in there to start like cutting it up and building it. But what we can do now, because these various AI agents are so good with code, describe what we want in plain English. And AI will generate a fully clickable prototype in just a few minutes. If you use Cloud Code, there's a front end design skill that really just ramps this up. It's really cool. I can't brag on it enough. But it will build out these beautiful mock ups for websites just with us going back and forth a little bit. And we can hand those HTML files to stakeholders. They click through what feels like a real page versus a static mock up. And this feedback is instant and specific. Move this button around. Add a filter here. This flow doesn't make sense. The mock up costs almost nothing to produce and we can throw it away and start over without losing weeks of work. Whereas before with the designer that would have been tragic. And without the designer we would have coded all that flow by hand and done like actual like implementation work. But with a prototype, it's very quick. It's very easy. The prototype's approved. The same code becomes the starting point for real development. It's very quick and easy and it really gives us a better starting point, whether that's building a dashboard or building a whole website or an app. A really interesting example of this is using that to pass along the information as well. InnerWorks does a lot of work with BI software. And so don't get me wrong, there's some amazing software out there, Sigma, Power BI, Tableau, that type of stuff that's still really, really good and valuable. But I have this friend, not at InnerWorks, who has really taken this code based approach to a next level. And he's not a business intelligence, like, software user. He's not a developer. He just kind of uses Quad to distribute information to his clients. And it's really small stuff. Of course, if it needed to be refreshed or updated, that type of stuff, he would probably use a real system for it. But instead of relying on those expensive BI tools, he describes what he needs to quad code and generates a single page HTML app with charts and graphs and filters. And with a few minutes, a little bit of validation work, it's ready to open in any browser. And that's something that he can hand out to his clients. At the end of the day, is it production ready? Is it going to be easy to update? Is it going to have bugs? Like, it's going to have issues. But if it needs a tweak, he can just throw that information in quad code. He doesn't have to file a ticket with the developer team. He just kind of updates it. And for his particular use case, that works really well. It's something that's small and it's lightweight, I mean he doesn't need to use a developer and maintain it. He's a small shop and so he doesn't need to outsource that information. It's just something that works for his flow. Software cost went from, you know, thousands of dollars in software licenses to nearly zero, whatever Quad cost him. Now, of course, there's ways that you want to productionalize this and that type of stuff. And so it's not a flow that works for everybody. It's a really good example of how Quad can really help our day to day, how AI can make something like this more approachable than it ever has been. Another really interesting example of this is just having a one page brief HTML that is Claude working to do that research as well. Anytime I need to organize information, compare opinions or options, prepare for a conversation, I love to generate a single page HTML file that just has all of that information. AI is really good about giving you tons and tons of information, but if it outputs it to something like an HTML file, it's very easy to see visually. It's kind of like the difference between rows of data versus a chart built in Tableau. It's just easier to approach. And so a couple of examples of this, I've had Quad look at comparing software products and to build a comparison matrix, something that's filterable, color coded, shareable with my teammates. Another kind of silly one, we at EnerWorks switched health insurance companies not too long ago, and I was like, I don't know the difference between these. And so I fed the plan PDFs to Quad, and I asked for a side by side like breakdown and things that would change, cost projections, that type of stuff. And they did a great job building out that kind of comparison document. Another really interesting example, kind of like our conference research, you can do the same thing in meeting prep for a client. We could give Quad an agenda, an attendee list, context, maybe some background information, have it build out a one page brief with talking points and background on various people. Maybe even things like meeting recaps so we can have that clean summary as we're going into that meeting. Things like action items hold in automatically as well. Or even post the meeting something that you could hand over to other people with that based on the meeting transcript. It really only takes a few sentences of direction and some raw data. And Claude handles the research, the layout, the design. You open a browser and it looks really polished. This doesn't require a developer anymore. And at the end of the day, even though it is code, it's not really something that's going to be insecure either because it's literally a one page document that you're handing out. You don't have to worry about it like you would other things in the past. Now if you think about the final iteration of this, taking it to the next level, the really cool thing that you can do when you get all of this together is you can start building autonomous agents that really are doing a lot of these pieces on their own. At Interworks, we have a couple of these in the works, a couple of things that are going on behind the scenes. I'll run through them really fast because we are running a little close on time here. But one of my favorites is we have a self documenting code base. We talked a little bit about this where I get prompts developers to update documentation. But another flow that was kind of our first iteration of this was skills and quad where developers could go ask information. Like if there's a bug going on, it can look across developer documentation as well as external client facing stuff. If it can't find it, it actually looks across the code base as well and brings some information to the surface. And that prompts the developer to document that if it's something that's missing. It's something that lives in our repository. It's versioned with code, not a separate wiki that kind of drifts or changes. But if you think about this flow, it's something that could really generalize a lot of things. Compliance, legal questions, customer support, academic research, that type of stuff. Taking it to the next level though, if you go on the Curator website and you search our documentation and you don't find a good answer, we have this nightly pipeline that actually looks at that and says like, is this something that we want to document? Is this something that should be documented? Is it a gap in our documentation? And if it is, it actually automatically builds out some documentation around that, creates a new pull request for our developers to review, to look at, and say, yeah, this is good to go. Let's add this to the documentation. Of course, other areas you can do that would be like support knowledge bases, FAQs, IT runbooks, HR policy portals. There's a lot of different things that could use this flow. The other thing that people talk about is a software factory, if you will, where AI takes and processes a lot more than just documentation. And we have a couple of these flows as well. Basically, AI can find problems automatically when things surface in the log or logs. It can write fixes. It can run tests. It can basically prepare everything and get it ready for a human reviewer. We have a couple of these in place. There's one that keeps up with the documentation. We've talked about it a lot. One that fixes code when things fall over. One that monitors the other two. There's a lot of different kind of areas for this. But these run on a little server just as a cron in the background over and over and over again. They're very interesting and very fun to watch, and they just kind of autonomously do their thing, and they'll bring up a pull request every so often. Another really interesting example of this is updating dependencies and keeping them up to date. Because as you know, in the world of AI, things are going faster and quicker than ever before. And often things need updated to keep up with the security because things are being hacked and issues are popping up very quickly. And so that was becoming a big bottleneck, and we have an agent that takes care of a lot of that now as well. Go ahead and skip to the end of here and we'll talk about this last slide. So if you think about the different tiers of AI, we've talked about a lot of these today. But as you find yourself on your journey, your path might look very different. You might be at a tier two where you're just diving in. Or maybe you're further along, you're implementing code with AI. Whatever that happens to be, there's different ways to kind of get to that next step if you know what they are, what you're aiming for, what your goal is. At the end of the day, though, the pattern never changes. You're the domain expert. AI is this execution layer. Tier one, you're describing the problem. AI is solving it. But by the time you get to Tier five, you're more or less just kind of giving guidance. You ask questions. AI captures it. It organizes. It processes. What really changes is the sophistication of the handoff. But really, with each and every one of those steps, every rung is worth climbing and something that I would highly recommend as you dive into your organization, just moving forward in that flow. With that, though, I would love to just open up for questions if anybody has some quick things. Or feel free to reach out and we can talk about this more. In particular, I know there's some really cool stuff on those upper end tiers that I would love to show off and then dive into more if we had enough time. But without any questions, can either drop them in the chat or I think we can raise hands here too. It looks like there are some in the Q and A actually. So you should see an option on the bottom to hit more and click Q and A. We have a couple of questions, one from Beth Murdock and then one from Bill Jones in there as well. Awesome. I was looking in the wrong chat. Okay, so Beth says, I don't write code. One thing I don't understand is when I can use the GPT to do work and when I need to use GPT to write the code to do the work. How do we know when to do this? So that's a great question, Beth, and honestly one that I personally struggle with as well. The widening there can be really blurry these days. The thing that I like to do is if I can get AI to write code to do something and it's going to be something that happens again, I like to get it to write that code. And the reason is then I use less tokens in the future. So if you think about something like a big AI system that we have one that that looks at all of our code changes and says like, is this ready to ship? And it takes a lot of logic to do that. And the thing was burning a lot of tokens, a lot of money. And I I started looking at it and saying like, what pieces of this actually needs to be AI versus what can be deterministic? What what is something that's scriptable? And when we started scripting it, the thing went from, like, several dollars a day to like just a few pennies. And the reason was we were doing a lot of logic with AI that we didn't have to, that we could have scripted. And we can use AI to write those scripts, And it really makes that approachable in a way that's never been before versus having AI process all of it. Okay. Just browsing over a couple of these. Okay. So the next question here is about best practices around information storage when using AI. And there is a couple of different approaches to this. But basically, the question is where should the data be stored? You can with the chatbot, for instance, you can upload a lot of information in there, or you can keep things on your desktop and and run like a desktop processing agent or or somewhere in between. Right? You could have AI reach out through, they call them MCPs, reach out to things like Box or Google Drive, that type of stuff and pull that information. And really, there's a lot of different ways to approach it. I don't think you're gonna go wrong either way, but if you can do local stuff, local processing, it'll be faster and easier. For instance, I use quad code, which is like actually running on my machine versus using like the chat app directly on Anthropix website a lot of times just because it's going to be quicker to access the files that are in my code base. And that could be it doesn't have to be code. It could be marked up files, PDFs, Excel files. The same is kinda true for all of them. It's gonna be faster there if you can do it there. But that being said, it should work just as well directly in the chat app for most of that depending on the complexity. When it gets more complex, local's definitely gonna be better, though. Yeah, so just a comment from Gordon. He mentioned that his team uses something called Fathom dot video for recording meeting notes and transcripts, and that has been very, very helpful. I would say, yeah, that's there's various tools out there that are so, so good with that. Zoom has some built in. Various things like Google Meet do as well, and they and they're varying levels of of good. Play around with what works for you. But there are some fantastic tools out there. I think that will be something over the next couple of years that just gets really, really exciting. Then we have one from Dan. Dan asked, how do you future proof your AI agents in case your organization decides to move from quad code to Copilot, for example? So there's various approaches to this at a very base level. The instruction set that I use for Quad could very easily be ported to CoPilot just by telling one or the other what I'm doing. And it can kinda do that migration for you. And so I wouldn't overthink it necessarily because you can Quad uses something called a QuadMD file and some various hooks and that type of stuff. And there might be various ways that your setup works, but it can also transform that to work with OpenAI infrastructure pretty easily. I've actually done that a couple of times in various projects and it works very successfully. And so future proofing didn't need to happen as much there as you would expect. With that being said, there are a lot of kind of open standards, if you will, that make that a lot more approachable. In particular, the instruction file, the big one that most people use in the quad environment is called a quad. Md file, whereas other AI systems like OpenAI, they'll use something else. But there's a standard one called Agents. Md, and if you use that, Quad will actually pick that up as well as the OpenAI infrastructure like Codex. And so that's one way to really approach that. But that being said, in a lot of ways, they kind of work with both. When you start getting into the agentic world and the agent's running, you you might need to do some tweaks. But really, I think AI can help you get there as well. When you think about developing the agents, honestly, wouldn't overcomplicate that even. If you go into quad code or codex or whatever and you tell it you wanna build out an agent, it can help you with like ninety percent of that as well. And so a lot of that is kind of the AI processing side in helping you with those those things. Okay. One more here. Do you recommend keeping a separate environment, GitHub, etcetera, for the AI coders as an air gap? Yes. So as a as a developer, I absolutely would not do much without GitHub or some kind of, like, versioning system behind the scenes. If you're just building out simple HTML files that are gonna live and die and and just, like, be very transact transactional, then you maybe don't. But if you're building out, like, some kind of application or something that's gonna stick around for a while, using some kind of versioning system is is very very important because you can insert in things like security scans or code quality scans, that type of stuff, and automatically spin it up on a server when it when it passes that stuff, when it has human approvals. Very, very important, especially in the world of AI because you never know, like, what's going to what's going to happen there. Right? Like, you hear these stories about AI dropping people's production databases. It's because they don't have that kind of separation in there. In my dev environment, that would never happen because I have the the dev database and the dev code, and it's all very localized to my machine. And then when things are ready to go to production, we use GitHub as that, like, intermediary layer where we commit the code and and push it up. And so it's gonna have all kinds of scans before it ever ever goes live. Alright. I think that's it. If there's any other questions, feel free to drop them in. Otherwise, feel free to reach out as well. We'd love to keep the conversation going.