How's it going? Yeah, I got it, and you? I can't hear you. Can you hear me? I can hear you. Give me a sec. Can you hear me now? Oh, there we go. You got me. Okay, cool. Perfect. Oh, there we go. Okay, cool. Sorry, just an audio glitch. You've always got to have something in any virtual meeting. There's always something. It's like 2021 all over again. While we're getting into it, I'm curious, man. I can see you, obviously, a huge AI aficionado. How did you get into AI? I'm always curious about people's stories and what switched them on to this stuff. Well, the short answer is curiosity, but, I guess, you've got to answer where the curiosity comes from. I think the way I think, that's quite nicely the way I'm seeing It things. So, having structured thoughts and then being able to access it. So, it's not relying on memory, it's relying on retrieval. And I just find the way that It thinks and then can produce stuff. It just seems to map with how I like to think. Yeah, I've got exactly the same feeling. It's kind of like, you know, it's sort of like this stuff that you've always wanted to do, but there was too much tedium involved in working out every single detail. And now it's like this unlock where you can suddenly explore so many things that you've never imagined or so many intentions. Yeah, absolutely. Because I hate tedious, monotonous work. I just can't. But a huge portion of our business is that. You know, we're basically like an administrative function for clients, but they don't have to deal with that. But we've got so many ideas, I've got so many ideas, and now it's just collapsing those ideas. So, in the past, it might have been like three months and I need a team of people. Now it can be, I can prototype it and see if it works. I can get it to 70 percent. And if it works, okay, then to get from 70 to 100 is hard. But I can see it. Yeah, yeah, for sure. And that's the speed. Yeah, that's brilliant. And how did you get started? Were you just sort of chatting on ChatGPT at the beginning and then it just sort of snowballed from there? Or what was the kind of path for you? Yeah, so like you play around with it. You know, when they first came out, it was like, cute, you know, you can do like essays, you can do like, you know, I can do like an e-book and make me sound smart. And then you just start working. And a big difference came when 4.6 came. I'd be working, but it was a lot of back and forth and just didn't get a lot of the formatting. And it was just the output was great, but it just took a long time. Yeah. And then one day, I think it was like a Friday, I was working and I had an interview with Giles and I said, you know, I can feel something's different. And I didn't know there was a new model release. Amazing. And then obviously that day I saw like, and so 4.6 was like a big unlock for me. Yeah, for sure. How's it, Bronson? Hey, Bronson, how's it? How's it, guys? How are you doing? Yeah, good. Yeah, well, and yourself? And then. I'm sorry, I don't know what's going on with my signal. I'm just going to try to hop across here in a second. Cool. Okay, cool. Yeah. So you're going to get to that critical mass where the model is sort of strong enough to actually do what you intended as opposed to a million back and forth, right? Yeah. Yeah. Well, so it's a double-edged sword, right? So I've been reading a lot about Methos, this new one that's coming out. And it seems like we're going to be drip-fed the model because it's too powerful to be released to us common folks. For the peasants, right? We must just do it. Yeah. Exactly. So it's got to go through. Apparently, like the cybersecurity is a huge issue. So they're dealing with all the operating system companies first to get them to do it. And apparently they've been finding bugs that have been there for like 27 years, like critical. So we'll be drip-fed. But on the flip side, it's also like we spend a lot of work now building like the infrastructure and the harnesses, getting the things. But then three months later, the model comes out and none of that is needed. Yeah. Exactly. You can then just... So I'm always like in two months, like how do you think about doing so much work, getting like the project set up? And then Methos comes and it doesn't need as much context. It doesn't need as much. It's actually simpler. It's better. And now you've got to go and like re-org all the detail that like Sonic used to need. The Opus needed less. Methos is going to... So actually having less is actually more, unleashes more capability rather than too much detail. And then if it gets to a point where it's just as easy as a user using it, where does our value go versus like a client using it or another wealth management firm where they haven't had to architect and be thoughtful about the design? Yeah. That's true. So I guess the answer to that is then at that stage, there'll be a new unlock. They catch up and then there's a new unlock that you can then go and explore. But the catch up is real and fast. Exactly. Yeah. The catch up is brutal or thrilling, depending on which way you want to look at it. I experienced several, but yeah. Yeah. Exactly. Both. Yeah. My attitude towards it is like try and stay as close to the edge as you can and kind of catch the first wave earlier than the next guy. So like what we're doing now on this project, maybe in 18 months, you can voice note Methos, build me the system and it'll go, cool. I'm done. I'm like 15 minutes later, right? But the point is that we have built so much context and knowledge. So like I almost think of a lot of my projects as like preparing the ground for Methos to take it to the next level or whatever the next model is, because there's always going to be another one. There's always going to be another one after Methos, right? So if you get to the edge of what's possible today, which is pretty much what this project is, I'm pretty sure this will be one of those cutting edge systems around, right? In terms of using AI for a wealth manager. And then maybe in six months, Methos can do it like that. But, you know, we'll have moved on to the next thing and kind of be taking it to the next level. So you can only do what you can today, right? Until Anthropic unlock the next level for us. So, yeah. Yeah. It's just something to get my head around because in the past you work on something like meaningful, it lasts for a long period of time. You've got that. Whereas now it's like quicker. So I don't know. Sometimes it does put me a bit on edge. But at the same time, doing this project is actually helping us structure our data. Because our biggest issue or challenge is that the data is fragmented across Excel files, Cloud files, CRM, email. It's just structuring that all in one place. Because once you've got it, then it unlocks all of those capabilities. That's exactly it. Yeah. So I think like my whole approach to architecture for this project and everything else I do is like assume that the models are just going to get exponentially better because they are and then do the stuff that the model can't do. So no matter how good Methos is, it can't just magically get all of your data and make it structured, right? Yeah. You can do parts of that process, but you need to kind of like almost prepare the ground for it so that it can work and whatever model can work. And when you start doing stuff like, okay, we're going to build our own harness and we're going to use our own, like, you know, that stuff gets obliterated pretty quickly. And I've had this experience over the last 18 months working on this. Like you say, you do some amazing project or you come up with some amazing structure and then, you know, Opus 4.6 comes out and it's completely blown out of the water because it's irrelevant. You don't need it. The model just figures it out on its own. So I've kind of had this, I've been talked several times about this and I've learned my lesson. Just assume the models are going to get so much better and do the stuff that the model can never do, like giving access to the data, et cetera. So, yeah. Yeah, exactly. So just a quick question before we dive in. So when the data gets anonymized going into Claude, when we, in Claude, do we use the client's name? Because if we use the name, how does it know that name, the data? Yeah, so we'd have like a kind of a tokenization. So you'd have, you know, you're using the name, then like a sort of a code that it gets entered. So it's not Mr. Funamava, it's Client73. And then it gets translated back and forth as it's going in and out of Claude, yeah. Okay. So do we say Client73? Well, you use the name and Claude would do some mapping and exchange it back and forth, yeah. Okay. Okay. That's perfect. Okay, cool. Let me get into the meeting brief and then we'll get into it. And thanks for all the really detailed briefs and docs and whatever. It really helps. Yeah. Yeah. That's Claude translating just what my, well, our intention is, what we want. So I can understand it, but I can't write it. No, totally. I think everyone who's using AI is in exactly the same position at the moment, which is like you can express your thoughts and you, but you would never take the, like you wouldn't take the time to write that document. You could, right? If you put aside a day, but you don't have a day. So the only way it's going to happen is if you kind of roughly dictate it to Claude and get the output on the other side, right? Which makes sense. Yeah. Yeah, exactly. Okay. So to start with, I just want to work through. So the analogy that I've been doing is in Claude, I've got, let's just open it. I've got what I call the main workshop. So in this workshop, this is where it has the context of the whole playbook, the whole AI playbook. We've got 34 tasks that we've identified of where we would like AI or data structure to help us with work. Each time I do a task, I call it a garage project because I like analogy. So we've got the workshop. Then I work on a task, build it in here. When it's ready to go to its own project, it goes to its own project with its own files and instructions. And then that gets stored here because a lot of the tasks are interconnected and separate. So one feeds the other. So it needs to know what is in here and going across. So anything we want to know about this project, I just put in here. How do you get people in the organization following your exact structure that you set up here? Is that a hassle or are you finding that doable? No, we haven't really done that yet. So Bronson and Giles had access to the team plan, but I haven't shared any projects because I don't know what can be broken or where if something changes that I don't know. So for now, it's just like playing around with it. They can use theirs independently. But at the end of this project, then it'll be shared with the team. We'll have to do training. I'm running into a big problem at the moment with token usage because it's got so much context. So I'm hoping when there's more data in your area, I can thin my project files and context because when we get to the automated reporting, I'm on 100,000 tokens before we start, before I enter the first prompt, which is just becoming a big problem. No, for sure. And I see that across a lot of clients. So there's definitely good solutions to that, which we'll implement as part of the approach. So the secret is to keep all of this stuff in the database as opposed to flowing back and forth in the LLMs context window. So you only want to grab the relevant context from the database in a structured way when you need it, as opposed to jamming everything back and forth all the time. That's the short answer. Exactly, because one of the things I do want us to work through is I want to identify the files that I've got in here, which ones I can take out will be in the data layer, which ones will stay in the project and maybe can be made a bit more lean, and which ones can actually just be a skill that doesn't need to be either, because that's not clear to me. The other thing that is a typical issue here is auditability and going back. So say some sort of mistake happens, one of the admin team use a particular skill or kind of tool that you've built here, and then six weeks later, there's some sort of problem where wrong data was reported or whatever. It's pretty hard to audit in this framework, right? It's pretty hard to say exactly what was the calculation, where do we draw the number from? So that's the other reason why you want it all in the database, because then you record every single thing that goes back and forth, and you can say, on the 8th of April, this person drew this query, this is exactly the workflow that just followed, and you can just pull that all as a report, right? Whereas in this current framework, it's amazing for, as you say, kind of like workshopping stuff, and you have to have it, but it's really hard for running in production when there's risks and responsibilities and that sort of thing. So that's the point of the project. It's awesome that you guys have built this stuff, and it's just a matter of kind of hardening it in a way that's possible and repeatable, yeah. Yeah, yeah, exactly. Okay, so these are all the tasks. I don't know how to go about this, whether it would be better if I send it to you and then your clock can digest it, but in our mind, we broke it into three phases. So there were 10 tasks in phase one. Some are in progress, some are completed, some have not been started because it's reliant on other ones having been completed. So this was kind of our, I think, so it was like quick wins, then workflow integration against some start today, and then eventually leading to like automation. Yeah, got you. So that's the map for me is to say, okay, here's all the stuff that we want to do. Let's build that into the system. So I'm going to need to do all the detail work to translate that. So yeah, that's perfect. The fact that this is already, because in terms of the project, you were sort of talking about who's responsible where, you know, you guys have invented all this incredible demand-specific stuff that I basically just need to formalize. So I'm not inventing, for example, a portfolio of drift monitoring alerts. I mean, I came up with some stuff for the demo just for fun, but I'm not the expert. So, you know, I would take what you guys have invented, kind of harden and crystallize it, put it into a restructured form, and then you can use it. But I'm not inventing this stuff. So the fact that you already have is phenomenal, you know, so it's just my job to turn it into something consistent and solid that you can use reliably on an on-dem basis. Exactly. And that's exactly where the value has to lie, is in the domain knowledge, because everyone's got access to the same tools. So it's the knowledge of the context and the interconnectedness of all these transactions, clients, emails, our philosophy, all of that. In my mind, I want to go from the manual reporting to automated reporting to automated insight. That's like the next clear goal. Well, the next one is automated reporting. Let's get that done first. But then it's the insight where things are just being pushed and notified, and we're spending more time reviewing and deciding. I want to go from doing to deciding. Yeah. That's how I see it. Totally agree. So this is a map for me, so I kind of need to sort of process this. In addition to the spreadsheet, like, you obviously, you've got Python scripts, you've got all sorts of kind of like implementation details, markdown files, et cetera. Probably the easiest would be if I can get some sort of like shared folder or whatever you want to do in terms of sharing the data, where you can go and give me all of the raw materials, and I can just work through them systematically. That's probably step one for where we need to go. Okay. Okay, so let me see which ones to do. So I've got, so this is the main workshop. So here's where I put the, so these are current ones, and then these are previous ones, but I don't have previous ones with you. So it's just something like this, like sharing. Okay. Yeah, for sure. So those are the project files that get stored in the actual project. This is the tracker. Yeah. I put those in the previous one, V18. Baselines, so each task, it takes me through a series of questions to find out exactly what the tasks will entail. Yeah. So these are the ones I've done, so that would be quite handy. And then this is, I just use this when I'm building a new one, because it needs resources. Okay, so it's just sharing. One thing I noticed in your doc is you sort of said, like, we're going to kind of clean stuff up and whatever. So I'd kind of go with what you have right now, if you're okay with that, because it's going to create, like, a barrier where you're trying to clean stuff up, but rather, you know, delegate that to me. Let me see if I can figure it out. If I can't, I'll let you know. And if I find inconsistencies and errors, whatever, which they inevitably are going to be, then I'll come and ask you, how would you like me to resolve this? So it's kind of like, instead of it becoming a, you know, you've got a business to run, delegate it to me, let me figure it out. If I can't make sense of it, I'll come back with questions. But I'd rather kind of get going sooner on my side and, you know, push the complexity and the cleanup to me, since that's inevitably my job. And it'll help me to understand it, to work through it all. Okay. Yeah, that 100%, that'll help a lot. Because when I was doing this, I didn't think this was going to be the actual file that was going to be used for production. Right, yeah. No, of course. I mean, it just matters, right? I'm sure it's very bloated, and yeah, it's levels, and it's bloated, and it, you know, it needs a bit of trimming. Because, like, on the automated reporting one, the main one that we're focusing on, this is where I'm sorting 90,000 to 100,000 tokens. So it does it, and I don't have any space to do even, like, one tweak. Yeah, got you. Cool. Yeah, so I would much rather get, like, the raw material as it stands right now. And then it also gives, you know, it gives me the whole picture. And then I can come back to you and say, like, I love this, this we're going to have to maybe clean up a little bit, et cetera. So it's, you know, like, rather give me the full download, let me figure it out. Yes. And then I'll let you know where I run into hassles. Instead of you thinking, geez, I've got a firm to run, but also I need to clean this stuff up for Graham. That's going to just delay us and create a lot of work for you. Okay. Cool. Okay, cool. Okay, so it's definitely the main project. These are the tasks that are really prototypal in production. So I think these two folders, so the main one and the tasks. Yeah. These other ones, I've got a lot of, like, other side projects, but they're not really for, yeah, I think these are the main two. Perfect. In terms of the NDA, do you have an NDA that you can send? Yeah, I do. So, like, maybe while we're talking about that, in terms of, like, a sort of a formal agreement, I can also send you a draft of something. So we can have NDA plus accounts in terms of the commission. I mean, we've effectively agreed already, but it's nice to have something signed. So I'll send that to you as soon as possible. And then as soon as we've signed, we can get the data across. Yeah. Okay, perfect. It's just like for our compliance. No, of course. Yeah. Because a lot of these files, I think, have client-specific identifiers. For sure. Okay. So let me... Always a battle in sharing a screen to go and choose the exact file. There we go. Okay. Yeah. Yeah, the maintenance becomes tough. Yeah. Yeah, you went from having a job doing lots of manual admin to a job babysitting us. So your next job is going to be babysitting me while I babysit them. Let's see how that goes. Well, that's fine because each layer is a layer of abstraction. Exactly. So it's easy to just send Graham a voice note. Graham, come on, let's get this stuff going. Learn to set up a whole lot of Markdown files. Yeah, exactly. Yeah. Okay, so let me go back to the brief. So let me see. So that's the workshop and how that's structured. So I'll share that. Yeah. I don't think I need to go through two or three Garage projects with you because you'll see it in the... Let me read it. Yeah, it's going to make more sense to me if I can actually just go through all the detail and wrap my head around it. Yeah. Because I was going to do a live report generation, but one, my tokens are becoming like precious for me during each session and day and week. Yeah. And it's going to take a while. But I can show you what the output currently looks like. I'd love to see that. Yeah, I'd love to see the output. And if you can share the outputs with me in those folders that you share or in another folder, then it sort of gives me like a start with the end in mind, like this is what I need to be able to produce something like this. So if I've got the raw material and the final output, then I can figure out how to structure that process. Okay. So when I share it with you, because I'm going to share the main workshop and then there's full tasks. So when you open that up, these numbers align to what the actual task is on that project tracker. Okay. I see. Yeah. Perfect. Task four is the main one. This is the reporting. In here are the reports. So these are outputs that is delivered so far. Got it. Okay. Cool. Some of them are good, some of them are not. Is this stuff actually getting put in front of clients right now? Is it more that specific stuff or is it more experimental and like for your own internal use? No, I can't put this in front of clients yet. Cool. Okay. Yeah, it's not, it's not there. But this is what it, another thing I battle with is whenever I put a PNG logo in a project file and project knowledge, it strips the PNG and it puts a black background. So I always have to go and then put the bloody logos in so that there's a workaround around that. Yeah. We'll find one for sure. Yeah. It's amazing how it's such a consistent thing. Like you come up with this incredible process and take them out and you end up fiddling with like PNG backgrounds as opposed to like world-changing technology that you're working with. It's actually, I'm not even humiliated by it anymore. I just accept it as the way that technology works. But yeah. Yeah. It's very strange. It can help cure cancer, but it can't identify how many orbs I'm storing consistently. Yeah. Totally. Yeah. Exactly. But anyway. Okay. So, yeah. So this isn't really the final look and feel. This is just my first best attempt at getting what we want. Yeah. So it'll have like an exact summary with some narrative on there, which needs to be worked on. At the moment, it's just fairly factual stuff of what's happened. Okay. This is something we haven't discussed with you, but we've got like a subcompany in Legacy, which is called Marlowe Capital. Right. That's the company that actually runs the portfolios. Okay. So we've got different solutions. One of our main strategic goals is to get more client funds into our solutions. Right. Either in our solutions or just in like, you know, other funds and over time we want to migrate into ours. So the first step is making it more visible to the client. So anytime we do report on something, I always want it in there. So whichever one, so this client has these specific ones. So it only applies to the ones that the client is in. I've got a separate spreadsheet that's got all the data that you would need for. Gotcha. So it's got returns, benchmarks, the commentary. Got it. So this is a big upsell opportunity and a big commercial opportunity for the firm, right? It's massive. Yeah. Okay. Cool. It's massive. Yeah. It's massive. Yeah. It's massive. Yeah. It's massive. Yeah. It's massive. Yeah. It's massive. Yeah. Okay. Cool. It's, it's, it's, it's significant for us. So I want to make sure. Yeah. So I saw what you said about upload. I think just a flat Excel spreadsheet is perfect and we can just, I'll build basically an upload tool for you. So that's the other thing for the client. So I'll take the spreadsheet format that it's got, figure out a custom Python script that reads it and kind of transforms all of the data to get it into the database in a structured way. And all you have to do is take an Excel spreadsheet, drop it into a secure websites box and it will instantly get thrown into the database. So that's the simplest way to do it. So it's just, it's something I've done for plenty of clients and it works well. There's a little bit of fiddling at the beginning to get it perfect, but once it works, it just carries on working as long as you don't change the formats of the spreadsheet. So, yeah. Okay. So, so walk me through that. Okay. So I'll pull up this specific spreadsheet. So then you'll build an interface that we just drag and drop. Exactly. So it's like a webpage, you log in, username and password. There's a little box that says drop the latest Milo capital reports. You drop it. Yeah. In the background, a cloud machine. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. It picks out every number, in effect, and puts it into the database in a structured way. And then you're done. So basically converts the spreadsheet into structured data in your database. That's the, that's the idea. You think it's possible? Because the way I get this data in there, we've got fact sheets. So I upload fact sheets important and it redoes, because obviously these figures will change depending on the new month. So we just, maybe we just go straight from fact sheet to database and cut out the moment. If we can go to fact sheet to database, but the database knows, because when we do reports, having this, it knows how to put it. So when the, when there's a tool call, it can easily pull it. So yeah, so if we can upload the fact sheets and then it like maintains something like this, because I'd also don't want it reading like PDFs because I've realized it's much better at getting data like this than giving it a PDF. Yeah. So let me have a look at it. So if you give me the spreadsheet as it is, and then the source PDFs as well, I can see how doable that is. It kind of depends on the PDF. So again, boring technical stuff like, like PNGs, but some spreadsheets are quite easy to pause and like kind of read. And some actually turn out to be super tricky and you introduce a lot of error. So I just need to have a look at them and see how, how easy it's going to be. It's not that it's impossible. It's just from a scope point of view. I just need to see how tricky it can be. Cause like the thing you said about can't map a PNG, but can cure cancer. It's like there's some stuff that you can do so quickly and some stuff that becomes incredibly laborious. And, and PDF conversion is one of those things that sounds easy, but it's actually super laborious. So I just need to have a look at that. Yeah. Okay. If it can be done, cool. If not, then. Let me look at it. I'd love to. I'd love to, to do it straight from the source. Yeah. Sure. Yeah. Sorry, Graham. Is it possible? So we have like, within each of these AMCs, we've got two that sit with this particular company. We've got another one that sits with a company called Gen2, but basically like on a daily basis. Those portals have got all of this information in terms of like top holdings, daily map, daily performance. So, so you, is it not possible, and I don't know how easy or not easy it is to just get a direct feed from that on a daily basis, as opposed to like a monthly static sheet. Yeah. In principle, totally possible. It'll depend how those portals work. If they have some sort of programmatic access, like an API, we can figure out how to plug in. If they have like a standard, like CSV or Excel download, we could just get an administrator to go and push the button and download them at a standard time and again, drop them into like a web box where it gets uploaded. So I think we're going to have to kind of walk through it to say where's the right point to grab this. You always want to go as close to the source as possible. So if they've got, if they've got an API that we can plug into, that's the best, because then it's just machine to machine. We don't worry about it. If they've only got a spreadsheet download, then we'd have to allocate someone to download on a regular basis, throw it into our tool, like a web-based tool to extract from the spreadsheet, or we go further down to the PDFs or your spreadsheet. But yeah, the closer we can get to the source of the raw data, the more efficient and easy it's going to be to run it. Okay. Cool. So as I was looking through, they're not all with one custodian, because then it would be a lot easier, but we've kind of got four different custodians in both. Okay. Okay. Let me, yeah, let me have a look. I know that from all of them, you can pull a daily CSV that's got like all your top holding performance, all of that. Let's start with that. I doubt they've got an easy to use API if they're not like huge companies. I don't know how, you know, it's not that common. Yeah, they are big companies. It's big Swiss custodians. Okay. Even if you take like Intuit or QuickBooks, right, they've got an API, but it's a nightmare. For example, like Xero is amazing. The API is brilliant. So you just never know if the API, so it's maybe for now, good to start with just the CSV download. We can get that working and then we can, we can kind of move upstream to the API as and when, you know. Perfect. Cool. Yeah. Just, yeah. Cause I suppose then it kind of just avoids, cause what we've seen is like, as soon as you dedicate a person to a task, there are delays. Yeah. It's like, oh shit. Sorry, I forgot. I was very busy yesterday. I didn't drop the file in. Where possible you want to eliminate human interaction. Yeah. Yeah. Cool. Let's do it in two steps. Let's get the CSVs in and figure out the data mapping stuff and just make sure that we can actually flow the data. And then once that's working with us, the two, let's go and figure out how the API works and just automate it. So we don't have to worry about an administrator being off sick or, or whatever. Cool. Cool. Yeah. That's a good point. So Bronson, can you download the CSVs for all of them and send it to Graham just as a one off so we can see, because that'll be part of this, because what goes into this as well is when we have investment committee meetings with each of these managers, all of that gets synthesized in here. Cause I want to know all of this stuff, where they sit on all of these and comments and what was action. So there's the data metrics, which we get from the CSVs or the API, but then there's also the qualitative management elements, which is this stuff. So that wouldn't replace us completely, but it could feed it. Right. Right. Yeah. Yeah. Yeah. Like some, some of the custodians, like UBS, where you allocate new capital, it just gets allocated according to existing holdings. But then with Maverix, which is another one that houses two of our solutions, where you allocate, it gets allocated to cash and like, we're not monitoring this on a daily basis, but I think what happens often is there's quite a large cash drag from like the moment the minute funds are invested to when the minute funds are allocated or invested to when they're allocated. And there is this delay in that, in that process. I think like for us when we have our IC meetings, if we can just sort of track, you know, kind of day-to-day or week-to-week, what was the cash holding and, you know, like just, it will keep us more in tune with these portfolios, which for us is vitally important. Yeah. Yeah. Okay. That makes a hell of a lot of sense. Yeah. I mean, what, what I'm sort of imagining there is like a, like a briefing from Claude as you're going in to say, hey Claude, give me a quick update on the cash drag and for this particular custodian, and you immediately get all of the facts. So cash was allocated on this day, this is the amount of cash drag, this is the impact on performance. And it's, it's, it's, it's brief already pre-packaged for you before you go into the meeting. So you can, you can push it. So it's such an interesting trend across all my clients is people using these tools to hold their service providers and partners accountable. So like you'll have a firm and I'll say like, we've got it, we've got a marketing agency and they're telling us X. But now we went and plugged in with our Claude to the actual data and we see why. So now we're going to have a chat with them. So it's so interesting because a lot of the time the barrier is not that you're not concerned about it. So like Bronson, you're saying, okay, what about the cash drag? What's going on here? But just the amount of hassle to get the facts straight before the meeting is almost too much to actually get there. Because if you can just delegate that to AI, suddenly you've got a very different relationship with your partners and service providers, right? It can still be a kind of a positive relationship, but it's much more fact based. Yeah. It's more transparent. Yeah. And I think that, that, that exact thing is going to come for us as well. So like what are we doing to OneBase? I think OneBase, I've been surprised because I looked under the hood properly and I've been very surprised at the data quality. So I thought you were paying for it. What is happening? Yeah, for sure. And it's so interesting because if you think about how many clients they have, I mean, I don't know how many, there must be dozens of clients, right? And none of them, the fact that those issues exist means none of those clients went as deep as you did. Yeah. Right. And told them, because if someone had told them, you assume they would have fixed it by now. Right. So it just shows the sort of like trust, lack of transparency thing is so pervasive, right? And AI just changes that. Yeah. Or just knowing how much patchwork is being done, because we're still getting the end thing, but that patchwork is delaying the deliverable. Yeah. Yeah. Yeah. Yeah. And why we need to double check it so much. Yeah. But it's coming for us as well, because at some stage, clients are going to have the same kind of process. Yeah. To us. It's a good moment to thank you for finding the bug in my dashboard. Yeah. So you're right. That was a mistake. So there's just a simple area in the code. I was missing a line that said, so the logic was it gets to the level where the human has to confirm it. So there's like a seeming match. And then someone has to intervene and say, is this true or not? You have to exercise human judgment. And the code is just missing a line to check that. So it goes straight to certain instead of probable. So the explanation is very simple. I just made a mistake. And the wrong line in the code. So obviously, we're not going to shift that sort of thing. We're not going to do that. We're not going to do that. We're not going to do that. Yeah. But it is about getting a stack on my side. So thanks for finding it. And I'll obviously fix and test everything properly before we do anything in production for the firm. Okay. Cool. Yeah. But good find. Good find. Good find. Yeah. Well, it's just because I've seen the knock on the door and the person's coming in, and I've got to go in. And they're all looking at the door. And I'm like, I'm there. I'm there. And then they're like, I'm there. And the guy's like, I'm there. And I'm like, I'm there. And then they're like, I'm there. And then they're like, I'm there. And then the guy's like, I'm there. Okay. Yeah, so we'll do, I'll send you this as well. So you can get the data via APR or CSVs and then to have like a scorecard because we'll get to the dashboard element, but this would be a nice one for us to see because we want a rate all of our managers and these different criteria and so this sound like navigate over here or this stuff to hold it more countable. So the way that I would kind of visualize this happening, and it's similar to capturing the qualitative views of the firm. So it was one of your points on the briefing that you gave is like, how do we, how do we get the briefing kind of qualitative views into the system? And the way I see that is, is through conversation with Claude, basically. So take this example, you've got these different custodians and you would have all of that in the database. And then you as a team would sit down with Claude and say, okay, let's assess, get to a point, Claude would say, okay, so for Sean Prong, this is what we think, this is what we're going to say confirmed. You say confirmed, it gets loaded into the database, right? So it's through discussion, human discussion, debate back and forth, gets formalized, and then Claude submits it into the database once you've given your consent and said, yes, I agree with this. And that then becomes a hard part of the infrastructure. So, you know, so if you've got a view on the Iran war or you've got a view on how Stratbench is performing or whatever, all of that happens in discussion and debates, and that's your human creativity and intuition. And then Claude helps you to kind of like elaborate that and then says, hey, guys, this is the view. Do you want to lock this? You say, yes, lock it. It goes into the database next. So you're not sitting at your machine trying to type stuff out or whatever. You're involved in a creative discussion. And the administration of maintaining the formal records around all the things that you've decided, that's the only one central database. So whenever you want to pull that qualitative view, it gets allocated to the correct kind of context. So when you're producing a particular client report, you're producing feedback for a custodian or whatever, it all pulls from that one source of truth. That's the way I would map it to. I think I think that's that would be by far my preference, because that's what I was asking over here, because I think just the structure doesn't make as much sense because it's an iterative thing. And I want our process then captured in a way rather than the opposite way. So I just wanted to make sure that that kind of process would work. It will work and it's something I've actually done for other clients. So it works really well because people work at a sort of conversational debate level and that's where we like to be. But then the hassle of entering it all into a system actually overwhelms the creativity. So you get out of your flow. Right. And that's damaging. So you want to stay in flow and the AI is filing and and organizing in the background. Right. But obviously asking you for confirmation. So they'll say, can I file this particular entry against this category? Yes. Cool. Next item. So that's that's that works very well. Exactly. 100 percent, because that's the domain knowledge being captured rather than a generic entry. OK, perfect. OK, so I'll just run through this. This should probably cover exact summary, a bit of commentary. And then there's always the South African version and an awful version which we separate. So then there's all the SA accounts. So it's just like a summary for the month. Then from inception. So the current problem is iInception with OneBase is only from June 2023, but we've got accounts from way before that. So we're working with them to backfill the full history. And that's a separate thing. But for now, let's just work with. But it should just be as easy as uploading the back history. Yeah. But because the rules will be the same. Well, I assume that would be the same. OK, so but that rolling return period, then a dashboard. This is something I'm actually battling with because it doesn't do this. It generates an image and then puts it here. So I can't edit this. And I've tried to get better. And I've tried it so many times to not have this. Yeah. So yeah, that dashboard is like a super tricky thing or like, you know, because of all these sort of trivial things like the labels in the wrong place. So I use I use an open source package called evidence.dev, which is the Oak Bridge example is built in that. And it just solves all of these tiny little niggling, irritating things, you know, colors and labels and whatever. It's a little tricky to work with, but I've kind of gotten used to it over time. So my idea would be to build this into a live dashboard. And the PDF or the PowerPoint that you pull is just a view into the actual underlying dashboard. So that gets generated dynamically. And then it pulls a PDF or pulls a PowerPoint or whatever. So I think that sort of stuff is solvable. You might get too cocky about it, but it's something I've done a lot of work on because I also hate that sort of thing. Why is the text in the right place? Yeah. And also the scaling, because we put total, some accounts, if they're very small. Yeah. So I don't know if I've quite got the scaling logic right. OK. So if you do that, can I then put that dashboard image in like this PowerPoint so the PowerPoint can be sent? Yeah. So is it a live dashboard? So it can. It's a live dashboard and you can extract it into a PDF, which could be in a horizontal format like this. In terms of an actual PowerPoint that you can hand edit, I'm not sure. I haven't done that before, but let me let me figure it out. OK. OK. I think that this is an important point because our ideal output is actually like a link to HTML site. Yeah. That's easier. That's easier than a PDF or PowerPoint. So yeah. And that's why I prefer it, because the advantage of that is I can send instead of 12, what is this, 16 pages. Yeah. If you've got a dashboard, they can toggle in that. It's not overwhelming. All the dates are all there. They can go and play around. Exactly. But the problem. Yeah. OK. Because the problem we had was when we're working on personal code, we could share artifacts in team code. You can't share artifacts outside your organization. So then we had to like have the HTML file attached, which is not clean. So you want to give the person like a persistent place where they can see their stuff anytime and they can go and log in. So, you know, Graham is a client. He has a username and a password. He goes to his site. He logs in and he sees his latest numbers. He can fiddle with the graphs. He can download the numbers himself. He can do whatever. That's ideal, right? Yeah, that's absolutely ideal. Yeah. I can quickly. Sorry, I can show you like what I did for a client recently. I think to Don's point, the challenge that we've got is that I, in essence, need to update and download a new HTML every time. It's not like a live document that kind of as we update it, they keep the same link that they can keep clicking, if that makes sense. So you kind of need something that's like continuous as opposed to like every time we're sending them new stuff. But I can give you an idea. Cool. Is that possible, Graham? Why don't you bring it up for us? It's definitely possible. So you've clicked around the Oakbridge demo, right? So the way that that works is there's a database that feeds the demo and you kind of rebuild that site on an ongoing basis and it pulls whatever the latest data is. So anyone who accesses that links and goes and looks at Oakbridge can see those pages. And if you have a client, Graham Rowe, for Legacy Family Wealth, he would be able to go to a similar thing that looks like the Oakbridge thing, designed according to our intention, but that has his information and you'll have different tabs and pages you can look at and you can toggle the graphs like you've been toggling and download the tables. And it's a living map of his portfolio that gets updated as the data refreshes. So and password and access control. So that's definitely doable. It's not quite in the scope of what I'd thought about, but it's absolutely 100% doable at a tech level. Yeah, because if we could rather have like something interactive, like a dashboard as opposed to a PDF, I think to Don's point, like a PDF is still like a lot of information. So can you see my screen? Yeah, yeah, yeah, yeah. So basically, like if the client got this as a link, let's say before the meeting or just like once a month or that, I suppose they've got a link to the portal. So they've got a link and they just bookmark it in their web browser or on their phone. Every time they want to go in, they click the link and they can always see exactly what's going on in their portfolio. That's it's it's it's the same link and the same username and password. Yeah, perfect. If you go in here, you can then see like portfolio summary. So total net worth. This is like a report on everything for this particular client, like his whole portfolio. But you can see like across the different entities that he's got his personal, all the assets in his personal capacity, his local trust, his offshore trust, his offshore company, then like the actual portfolios we manage. So like if you go into here, you can see what is the total value, what is Marlowe make up of that, other local equity funds and what is his cash track in the Marlowe AMCs, just like what is the risk return for all of them. This is this is really so the one thing that I would I would sort of add with the evidence dashboard is like if you look at this table, we could put it in this balance. There'd be a little button in the bottom right. Like there's an Oakbridge thing where it's like download a spreadsheet. Right. So the client could actually get their own data and play with it. And the graphs would also be interactive. So as you highlight those different bars, it would sort of pop out and all that. Those are nice touches that give people a bit of a thrill, you know. So and they could download the whole thing as a PDF as well. So there'd be a little button top right that says download this PDF so they can go and keep it in that form. But all of this stuff is very doable as a living dashboard for each client. Yeah. As long as we got the data needs. Perfect. Yeah. So how do we solve this? No, I'm done with this is just to give you an idea. And this is the challenge for us, but every not for us, for me, the portal I've created for John is different to the portal I've created for John. It's different to what I created for Graham. So to manage these things is a hack because I'm like, shit, what did I put in John's one? And then when I go create a new dashboard, like the dollar sign maybe pulls through and it's got a K instead of like the full digits. It's like there's all these nuances. So it kind of needs to be standardized. That online portal also prompts clients to take like online meetings or meetings in our office where we've got a proper TV screen and it's like, you know, it's a bit more professional as opposed to like I hate meeting clients at a restaurant. It's just a disaster. You can't read upside down. So, yeah, like if there's a way that we could have a live portal like that for each client, but it's pulling through like the same look, the same standard on standardized and depending on what you want to show the client, which John's done very well, it's like basic or sophisticated. And then you can have like a quarterly report, which is like only on the investments we manage and maybe something that's like a little bit more comprehensive. It's like an annual family meeting. So, yeah, there's just at the moment every meeting I have, I have to download HTML and I share that and the client asks if I can send it to them after the meeting. But then like I know that that info is going to be redundant in the future. So in this case, you just say, look at your report. And the cool thing is that as you innovate, like say you go to one particular client meeting and the guy says, I want to know this for the annual family meeting. And you say, oh shit, we don't have that. You bake it in, then it gets automatically deployed to every single client. So the next meeting you have with Graham, you say, hey, by the way, you didn't ask for it, but we've got this cool new functionality. Go and open your site. Look at the bottom there. There's this new tab. Isn't that cool? So, you know, you can also use that commercially to drive people to Marlowe, for example. You know, so it's all about just kind of like creating that visibility for every client. But yeah, it's like a massive hack to manage it, like in multiple instances. So we'd have one way of doing it, which you can maybe toggle different levels of sophistication depending on the client, but it would be a dynamically refreshed view that people can log into. Look, I mean, it's not part of the code, Scott, but it's definitely doable. I don't know if we want to do that as another phase or I need to think about it, but can it be done? Yes, definitely. Yeah, OK, because, OK, so would everyone then get the same dashboard, so it'd be the same. Exactly. So there'd be one LFW dashboard format, which you're continuously improving. Every client has access to the same one. And the way you solve that, like, what do they want, complexities, you maybe have some toggles that say, OK, here's the basic version, here's the, I don't know, we can come up with nice names for it, but here's the deep dive version, whatever. So everyone's getting exactly the same thing, but you sort of you can you can go into different levels of depth depending on what you need for that client, if they're brand new, if they've been around for ages, et cetera. But the point is, then there's nothing to maintain. Both you and the client just click on that link and it's there, right? It's not about manually fiddling with anything. I think initially, I think I would like us as wealth managers to be using that and then only deploy it to clients once we've used it for quite a while. Once you come to rather than. The other cool thing is, I mean, you guys obviously already doing this, but when the data is in the database, you can also connect Claude via an MCP connected to that dashboard. Right. So you can say, OK, I'm getting ready for my meeting with Graham. Have a look at Graham's dashboard. Let's have a conversation. What do you think we should chat about? I know he was worried about Trump and Claude will then read what's on the dashboard. And second, given Graham's crypto holdings and his concentration and energy, X, Y, Z. Right. So the AI becomes an ally in reading that particular dashboard against their clients and prepping for the meeting. So that's another thing that's quite and that doesn't have to be a full market. It's more just like someone advising you in the lead up to the meeting on relevant things to talk about based on your context about that client. But it's brilliant that you've got all of this stuff prototyped and drafted already because it makes my job super easy. I just need to kind of formalize it. I don't need to invent anything. Yeah. When you say all of it, I think you're we've got we've got a good base. Yeah, we've got a good base. Like what I'm seeing there is brilliant stuff. So it's and the point is, it's not theoretical. It's informed by your actual practical experience of talking to clients and getting them to understand what's going on. Right. So that's what's so valuable about it. Exactly, because our value is going to be what questions do we ask or what do we want to extract? What way do we want to take the conversation? But it can come up with ideas. Look, I want you to rebalance or deploy cash. Give me ideas. Yes. I want to go from like doing again, like deciding, OK, well, let's do that. Follow more of that thread. And helping with sales as well. Right. So you say, like, what would you say to a client to get them allocating more to Marlowe? Fact based. True. What have we got? And then you've got some advice and bullet points to talk through instead of you having to burn your own mental energy going looking at stuff, you just give me the ideas you verify quickly. And now you've got you kind of armed with that knowledge going into the meeting. Yeah. And a big part for us is personalizing the context, content to the context. Yes. So that if it's connected to the email, so it also knows how you interact with this client, open items you've been discussing, its portfolio history, when things were done, that that is the true value add. That's how we position a portfolio because we show how that portfolio fits their goals, what they're trying to do. This is just the way to do it. But it's not that this is amazing. It's just like this is just a natural thing to be incorporated. I don't want to focus on like the product features. It's more just like what is this doing for you? Yeah. Got you. Makes perfect sense. OK. OK. So, OK, let me just finish this round. I start playing a video game. Trying to find the right way to play the game now. I mean, because I don't know if you're going to work because I go there and then I've got to try and quickly. Oh my goodness. All good, man. This is painful. You're welcome to kill the screen share while you're finding stuff. Yeah. I just put up a clue. Yeah. Yeah. So Graham, on the Marlowe Capital side, I suppose we're constantly asking ourselves. So it sounds ridiculous, but to try and get like a snapshot. So our best ideas portfolio benchmark is 65 MSCI Equi and 35 Bloomberg Global Ag, which is basically like 65 equity, 35 fixed income. And that's a live trackable benchmark. But like when we have these meetings, it could be great to look just on a daily or weekly basis. You can see exactly like how you're tracking against benchmark. And then in that peer group, they are like tens of thousands of funds, ETFs, you name it. And it's not really to go and scan us against like every single one of those, although you probably could. But really to like flag, like we've got one active fixed income manager that like really underperformed last year. We've got a good relationship with them. So we've kept it in the portfolio, but they're like on our watch list, if I could call it that. So like maybe some managers that are on your watch list, but you're constantly there, like really are putting them up against like their benchmark. So like their benchmark would be the US high yield. And they're just tagging things like costs, positioning, kind of like so you know that we are way more informed. Because what ends up happening with an IC meeting is like, and Don, you mentioned it yesterday. We, it's always like, OK, cool, let's see how the next quarter goes. OK, cool. Like, yeah, we'll see in three months time. But you know, like a lot happens in three months. And I think if we can go into those meetings like super prepped to say like, guys, you know, there's four other managers that are active, same fee structure as yours. And they've all outperformed you by like 2 or 3 percent. For that reason, we're going to like start transitioning out. Just give us like maybe like a bit more like AI driven advice as opposed to us, which we can then use for those meetings. Yeah. So the way I would think about doing that is kind of like a defined workflow in Claude that lives actually in the database. So you would have this data about performance of different providers and fee structures. We'd have to put that into the database itself. And then you set up Claude, which is connected via NCP to the database. And you say, we're going to talk to this provider, run the whatever workflow, we can give it a fancy name, whatever IC prep workflow. And Claude then goes and uses whatever tools we built for it to pull the relevant information about that particular provider out of the database. And then it comes back to you and says, hey, Bronson, I've reviewed the last quarter. This is their performance versus benchmarks. This is what we need to discuss with them. These are the other guys who've outperformed them. Here's your kind of meeting brief. So it's like a dynamic meeting brief that gets created according to a predefined workflow. So the work is to get the data into the database and then to define the workflow that says when I'm going into an IC meeting, I want you to pull this data, that data and that data. And I want you to create the following summary and then I want you to brief me like this. So we define that once in the database. And every time you go to have an IC meeting, you say, hey, Claude, create the IC prep for X and thumbs up. A few minutes later, you've got that custom brief. So instead of having that as like a dashboard, because eventually you just have like too many dashboards, it's almost like a skill. It's like a skill, but it's more formalized in that it's baked into the actual core database itself. So that's the way that I would tackle that particular. So then the benefit of that is you're not maintaining the skills at like a project level in your local view. And Claude, the stuff lives like in a cloud location and it's a one formalized version. And then say you go to that IC meeting and you go, OK, well, that went well, but if I wish I'd been briefed on X, then you say, OK, cool, let's upgrade the workflow. So it always takes that into account and brings that in. And then that rolls out to every single IC brief that you have there after. So it's a way of also iterating your own IP around stuff. Does that make sense? Exactly. Yeah, it does. I've kind of got that mapped out pre-IC, after IC, post-IC, what should happen so we can work with what is currently there. So what we would do is bake this into a structured prompt for Claude in the database. But it'd be more than a prompt. It'd actually be like define the actual SQL script that you'd run. So pull this data from the database, turn it into this view, extract this number, put it all into a report for me. So that structure you just showed, like pre-generate this for every single IC meeting that I go to so that I've got a standard report that's up-to-date data every time I walk into a room. That's the idea, yeah. Yeah, exactly, because right now I've got it mapped, like in a document, that having it mapped in equivalence of a skill would be so much better because there are set things that need to be done during each of these phases. Exactly, so each of those we bake in as a, it's kind of a skill, but it's a more formal version of a skill. And you guys have already done a lot of work of figuring out what the process is as a matter of formalizing and putting it somewhere structured. And the cool thing about that is you can then iterate it as you learn. Yes. Cool. And then, so to Grant's point about putting like benchmark info, I've tried to set up connectors, but I haven't been very successful. So like if we can, because I see there must be places that you can pull. Yeah, yeah, there's going to be APIs that we can pull that benchmark data from, whether it's subscription, et cetera, we can figure that out when we get there. But the way that would work is we'd build an API connector into whatever provider gives us benchmark information. It might be even available under whatever existing subscriptions you have with data providers, or we might have to upgrade something. And then that would get collected by a worker that pulls from the API, like a cloud worker, and it goes and pulls on whatever frequency you want, daily or weekly or whatever, and then loads that into the database. So when you need access to information, it's already in your system. It's not something you have to go and fetch at the time, right? Yeah, because it might surprise you, but it's actually quite difficult just to get that data. Just so like, what did the JCL share do in March? What did it do over the last three months? It's like, everyone reports on it slightly differently, because Yahoo Finance has this stuff, so I don't know why we can't just pull from here. Morningstar is one of our DFMs, so surely we can, I tried to connect to their connector, but it didn't allow it. So I've actually got a Yahoo Finance connector already. You're probably going to laugh, but I built it for myself, for my own equity portfolio. So I track about 8,000 global stocks, all through Yahoo Finance. So I've got a database, I pull it all in, I pull in other benchmarking data, and then I've got everything set up in Cloud so I can chat to it. So I'll go and say, okay, how's the latest Trump shenanigans affecting the portfolio? Pulls all the stuff from Yahoo Finance. I've already built that as a hobby project for myself, so I understand how this sort of data pool works quite well, yeah. Oh, brilliant, yeah, because yes, that would help a lot. It's always rewarding when something we did that seems a bit mad, it turns out to be commercially useful. So that gives me a bit of a thrill. Yeah. I mean, who doesn't? No, our challenge, I think, with these IC meetings, sorry, we're veering away from what Don and I put together here, but the IC meetings is like, we're constantly questioning, like, what else is out there? How are we tracking against benchmark? Is it normal? Isn't it normal? Should we be looking at other managers? But because you don't have like a centralized database of, and Yahoo Finance is great for like individual equities, benchmarks, and ETFs, but manage active managers, like, might be a little bit trickier to get that data from there. But I suppose, I mean, we can figure that out. Yeah. So look, if we can get the data, so the stuff I've done is all at an individual stock level. So in terms of, you know, manager performance, I actually haven't done that. We'd have to figure that out. But what you could have is some sort of internal dashboard for yourself that tracks every single manager. And we have certain quality metrics that we want to see, you know, whatever, sharp, et cetera, et cetera. And then we rank order all of them and say, then when you go in to talk to one of your guys, the current provider, you say, we've compared you against the best in the world. This is where you rank. You're in this quartile. This is what the best guys in the world are doing. This is the fee structure. Let's talk. You know, so the hard part about that is just doing all the analysis up front. But if you can automate that, then you're in a much stronger position. Yeah. See, we can connect to like Morningstar, because they'll have fund information, not just stock or ETF. Yeah, yeah. Or S&P Global, I don't know what they've got, something like that. So, which subscriptions do you guys use at the moment to connect to that sort of data? OneBase have got, we've got a license for you, but that's where they currently pull all of that info, yeah. Okay, so it actually comes via OneBase, which is great. Well, yeah, like we have to pay for the subscription, because I don't think you can buy one subscription and run it out across multiple clients. I think that we've got our own subscription that is housed. Yeah. But the data is in your OneBase extracts at the moment? Yeah, some of it. Not really like the Marla-related data. The client-related data is, but then the Marla-related data isn't. Okay, cool. Look, I would imagine that a lot of it can, you know, a decent chunk of it can be done around benchmarking from open sources. But anyway, let's investigate that. Yeah, cool. It's basically, it's not can you do it, it's just how much will you have to pay for the data subscriptions to get access to this stuff in a programmatic way. So that's the question, yeah. Cool. Okay, okay, cool. Let's just finish working around here, then the asset classes. But like the asset classes, we've had to do our own asset class mapping, because different platforms classify assets differently. One will be equity, one will be alternative, you know, so we need a standardized thing. So we've got that mapped. Internally, I think we still need to fine-tune that, but it's nice to get that. Yeah. And then it's the same thing on the offshore side. So it just goes through the exact same thing, and then Bob's your uncle. The main thing is testing every single platform, like here's the logo that it just does it. Because the formatting, like yeah, it becomes a problem. When there's like the total is big, but there's lots of like smaller kind of like... Yeah, so you need like a log scale or an index scale or whatever, so... Yeah, and it needs to be like dynamic, and like, yeah, there's too many accounts, so then this goes, you know what I mean? So like... It kind of becomes unplayable in this format, yeah. So the benefit of the dashboard thing is you kind of solve it once. Like there are lots of fiddly things, but the point is you only have to solve it once, and then you roll it out to every single client, right? So it's not that they won't be teething problems, but they don't need to be solved over and over again. Yeah. Yeah. Okay. Okay, okay, perfect. Then should we move to the dashboard? Yeah. Just what our ideas are. So is there anything in here that was not like in the scope or that can't be done in phase one? I was very happy with your decisions around that, and yeah, I agree with the way that you laid it out. The model thing, obviously a little bit of a change, but happy to just fold it in, yeah. All good. Okay, perfect. So is there anything I need to go through here? No, I've had a look through it, and it makes a lot of sense. I mean, it's just the next step of actually getting access to the detail, which we've already discussed around the folders, so that I can start actually mapping it out, yeah. Yeah, and then these UX, so default to my clients, because otherwise it's going to be very busy, especially when you're looking at like transactions and that. Happy with that, yeah. Yeah. Okay. And then the role-based thing, yes, 100%. So like as a filter, that's perfect. Like actual like proper formal RBI AC is like a little bit more tricky. So as long as it's like it's a UX filter, but not like a hard kind of like data level filter, then it's pretty easy to do, and that's sort of fine within the existing scope, yeah. So everything can be shared by anyone who wants to look it up. I just want the default to be on one page. Because if you said to me, I can't let Bronson see Don's stuff ever, then that's a completely different proposition. So that's not what you said. Yeah. Okay, cool. Perfect. Okay, except that might apply though for like the fees, because for everyone in the firm to see everyone's fees, that may be tricky. Okay, cool. Let me figure that out, yeah. Let me have a look at that. Unless it's well hidden or I don't know, there's going to be some kind of barrier to just anyone just toggling there. Yeah, so we might put that behind a separate sort of access control portal. Let me think of the simplest way to do it. Because yeah, I agree, you don't really want to create consternation in the firm because people clicked on the wrong button, right? So yeah. Yeah. Okay. Okay, cool. Then for the full 34 tasks, like I was saying earlier, what is in your section, what's in our board and what can be a skill? Because I really want to clean up what goes in the projects. Yeah. And the skills, we've got some skills. I'd love to get like your take on like if there's obvious ones that should be there, review the existing ones to make sure like that. Yeah. Because I'm finding over time, maybe having a very defined skill can sometimes inhibit what the AI can actually produce because I've restricted it so much, you know, producing stuff. So over time, you know, I don't know how to play that. Yeah. So the interesting thing here is kind of like it's almost like this invention pipeline. Like you were talking about garage workshop factory maybe, right? So the skill is kind of like you can make it up easily. You can try it. You can experiment with it. Then you formalize it a little bit more. And eventually, once it's like we're sure this is something that we consistently want, then you put it as like a formal database structured workflow, right? So eventually, you want to pull as much stuff as possible into the database level. But you don't want to do that prematurely because then you're going to harden something that's actually broken or inappropriate or just doesn't really apply, right? So I'd see that partnership as you guys have the intuition, the knowledge, the experience, you're encountering clients every single day. You're inventing stuff at the skill level, like that wouldn't stop. But then there's almost a handoff to say, this is something that's getting boring. We're tired of babysitting it. It works perfectly. Graham, here you go. So that's the way that I would see that. Like ideally, as much of the firm's IP as possible is sitting at the database level, but there's always going to be a frontier of invention. Because basically, what's going to happen is the more efficiency you guys get, the more free time you're going to have. You're going to think of more ideas to make money or do more for clients, right? So it's never going to end that you're inventing new skills. It's just going to get more sophisticated and complex. So you can't say, put everything in the database. You can just say, stuff that we're really sure about gets like database level treatment, stuff that we're inventing sits at the skill level. And then we have a process for sort of porting things across continuously and building up the IP base. Yeah, I don't know if that makes sense as a philosophy. Yeah, so as an example, so like we've got legacy and we've got model capital. So these are just skills when the PowerPoint gets done. So would this be, once we're happy with what this produces, go to the database level? Yeah, exactly. And then you would, so what you would do is define all of the stuff very precisely and informally at the database level. And then when you're sitting in Claude, you would say, access the LFW PowerPoint workflow and create this. And then it'll go and fetch that formalized, finalized instruction and build you the PowerPoint locally. Okay, but would you have to invoke it? Like if I want the LFW? So you would just, you just need to, in plain English, you would just need to say, okay, this is the thing, do the LFW BBT thing. Or we could even hard-code it in the system prompt to say, just give me a second, you can keep the talk. Cool, perfect, see you now. Two minutes of a break. Sorry, nanny and baby working from home and just chatting to me, sorry. Okay, no worries. Okay, so yeah, so what we could do is say, we've got the PowerPoint skill at the database level embedded, and then we could put it in the project definition to say, whenever you're making a PowerPoint, use this particular workflow. So you don't have to repeat yourself, it's just, hey, I want the PowerPoint and it understands what you mean. So that's pretty easy to cover, yeah. Because at the moment, each project's got like defined outputs. I've got a brand router, so it then decides whether it's Marlowe or whether it's LFW, because it's different branding. Then it also asks my client's sophistication level, because I'm finding it very helpful to communicate to clients in a way that they understand. And having these different, you know, layman, informed client, sophisticated client, you know, can be simple, is a big thing, but at the moment, it just like prompts me, which level do you want? So you kind of want to keep the user deciding which one to do. Yeah, for sure. So it should, it'll be like that. It'll work exactly the same way, except it's just a little bit more formalized than the markdown folders, right? Yeah, exactly. So Claude would be able to figure it out. And we could have a mapping even in the database in terms of the client sophistication. So it would know, okay, Graham is at this level, let me automatically route over there for him, yeah. Cool. Okay, so that's something that learns, okay, this client, you've always selected this, so. Yeah. That then gets formalized. Defaulted, yeah. Or defaulted, yeah. Okay, okay. So how can we work through this to see what is? So I've had a look at the kind of choices that you made, and it seemed like a sensible set of choices. I think the only way for me to make more progress on it is to actually look in detail again at the actual docs. And then I'll say to you, okay, this one, you've said, Claude Native, that maybe it's a good candidate for database, or this skill should actually be here. You know, I'm gonna need to work through the details to get to a point where I can kind of make a final call. But when I looked at it in prep for the meeting, it all looked sensible from what I could see. Okay, okay, perfect. So I just got someone at the gate. Great, glad I'm not the only one with Claude Native. Yeah, we're all working from home. I don't know why I'm working from home, but my kids have gone back to school today, so it's actually quite relaxed. Oh, okay, how are the kids, eh? Oh, sweet, okay, nice. Yeah, my little girl's like six months old, so she's, yeah, she's pretty sweet. Oh, I hate to think, okay. Okay, so you'll be awake at one o'clock in the morning working on this, regardless of what time it is. I actually get a lot of work done, right? Because there isn't really much sleep, so I might as well, if she's not with me and she's with mom, I might as well be on my laptop. So yeah, that's what I'm doing. Well, Don, you may as well have a newborn, but I was with you at night as well, so. Yeah, yeah. Don, the kids yourself, how old are you? Four years old and 18 months. 18 months, okay, okay, cool. So you're also living in my sleep department. Yeah, he's going through thieving and now we have to do sleep training, so that's the new job. Sure, sure, sure. But it's all good. Yeah, no, it's all good, all good, perfect. Okay, so I just want to go, okay, so once you've got better or more insight into that, we can then see which one does not trust what I initially said there. I think first we already discussed. We spoke about this, yeah. So the source of truth will be our discussion, our inputs, and then that just gets captured at the page space in the way that you want it captured. On confirmation from you, yeah, yeah. Yeah, exactly. This is fine, we've discussed MOLO. Then for phase two of comm space is basically the one base, but for revenue. So well, that doesn't need to be done now. I'm still finding out how we can get those files, and then we can build a separate dashboard. There's lots we can do there. If we just have all the data, then we can like play around and see what, yeah. The flow verification, well spotted, like I said, yeah. I'll fix that. Just on the data that's going in, sorry, the data that we get from one base, I think as Don's mentioned, there's a bit of a challenge in terms of like the quality of the data at times, mainly just how they capture certain transactions. So like withdrawals versus rebalancing and a few other bits and pieces. I called yesterday and he took me through like some of the prep and the stuff that he's had to put together. And I suppose like the purpose of going through this exercise is to try to get like as efficient as possible without this like massive monthly task of like creating new sets of rules and rechecking dates. So what I was saying is like, if the data is 99% accurate, but to get it to 100% is going to take like double the amount of work. It's like, I suppose, we just want to figure out like what is the cleanest and simplest way to do this so that we are freeing up our time as opposed to like we've just employed a new lady that's going to be joining in May. And if she's spending like three days every month trying to like put data in and check accuracy and do all of that, like is that the best use of her time? Probably not. So again, like I suppose, you know, open to ideas once we get this process started, like if there's stuff that you see that you think we could maybe like tweak or change or you know, how do you restructure that? Like, please let us know. Yeah, no, with pleasure for sure. So yeah, look, Will, the idea here is very much to say, look, you guys have really thought through a lot of the logic of the process. My job is to kind of automate it so that instead of taking three days, it's clicking through a dashboard for an hour or two. You know, ideally, that's the idea. So it's to take all of the thinking and the processes that you've invented and the recons and then structure that so that it happens more automatically and it's a tool that your team can use to chart through this stuff. I think that Bottleneck pretty quickly is gonna move more and more to the OneBase guys because you're gonna be churning out mistakes feedback to them. So they're gonna have to up their game. It's already happened. Yeah, no, for sure. And if it's already happened without the tooling, imagine if it doesn't require you, Don, it just requires one of your administrators to click buttons to produce mistakes findings that go out to them. They're gonna be a bit, but look, it's a real favor to them. It's helping them improve their business, right? So they can't complain. And I told them that. Yeah, exactly. I think it's a massive opportunity for them. And it astounds me how they can't. So the way they do it is, they've got like LFW Glacier platform rules. Then another firm that another customer is their Glacier rules. They don't have a Glacier rules that goes on. And it was like, that just astonished me. So why? Maybe I should go pitch them once I'm- I'm telling you, I think, and their data aggregation layer, they're selling clean aggregated data because at the moment that is complex. Like if they could just, there's so much they could do with that. Interesting. They could be doing this. Yeah, but they're not yet. Because they've got all the data, but they're not. So, and that's fine. I don't want them to do like, you know, a product that they can go sell to other lock managers, because for now this is armored. But okay, so now for the testing. So I'm testing each platform, but I'm running into constraints because I'm starting with 90,000 tokens out of 200. It's taking long. Because this project is doing so much of the heavy lifting. It's going through, it's got to go through all of this. And I'm not happy with, some platforms are fine, but some are still like edge cases. I need to figure out a way how to test the stuff quicker. Yeah. If that makes sense. So I don't know if like, we take it as is, put it in the database, and then I can free up this project's capacity and then test. Because lots of this stuff doesn't need to be in here. Yeah. It can be a database. Or whether you need me to test each platform. You know, again, I'd say that the idea would be just throw the stuff at me as it is. I'll figure out how to make sense on my side. And then we can sort of iteratively pull stuff out of your environment into the formalized structure environment. So, you know, it's, it'll be a process like, but yeah, hopefully we can free up some context soon because it's sort of slowing down your whole process. Okay. So I'll share the files as is. Your base could be able to replicate this or however you, and then you can let me know, okay, we need more testing here. Let's do this. Exactly. Because it's, yeah, it's just taking too long, but also in your database, it can go through those four layers. Yeah. One base can go through those four layers and then I can like do the testing. That would help immensely because at the moment it's been so much. Yeah. And I've had to try and break it up because when it goes to the pipeline or the enrichment or the report folder, it's like, well, yeah, it's revving too hot. Yeah. As you say, it's amazing that this isn't something that OneBase have done because it's kind of their core job, you would imagine, right? And if they haven't done it, it's really interesting that no one else has done it because it just seems like such a dire need, right? So the fact that you as a wealth management firm have to kind of invent this stuff yourself is pretty astonishing, right? If you kind of take a step back and... It is astonishing, it's frustrating, but it's also an opportunity. An opportunity is not going to last forever because at some stage, you know, all of these arbitrage gaps are going to be compressed. Yeah, for sure. Yeah, exactly. But for now, if we can get it right so that the whole point is that this delivers a better deliverable to the client. Yes, yeah. That is personalized, more updated. Eventually over time, I would love to just have the data just from the platforms come in. Yeah. That can go through the four layers, the same as OneBase is. How much they're actually cleaning up is now I'm questioning because there's 18% of transactions that are just not even ID. Yeah, yeah. It's like, how? But anyway, that's not now. Okay, well, that makes me happy that I can just share the files and then that can be... Yeah, just maybe you find yourself thinking, I need to clean this up first, just go, no, send it to Graph. Yeah. Okay, okay, perfect. Okay, so then that'll be like... So then the fundamental meeting brief. Yeah. So will that be like, that'll be a standard thing, call it a workflow that the database can handle. Then if we want individualized ones, so like sometimes I'll put like a different pack together, you know, attacking it from a different area. That will still be done at project level. Well, we can decide, you know, again, it's that flow of like what stays in the workshop and what goes into the factory. So you might say, hey, so this is our standard meeting brief, but actually there are three different flavors of meeting brief that we'd like to have access to. And we always do these three. Then we bake all three in at the factory level, right? And then invariably someone will get creative and create a fourth, and then maybe we leave that at a project level or whatever. So that's gonna be a big part of the success of this is managing. I can be innovating and coming up with cool ideas all the time. We're working with that at a skill and project level, but then we decide, okay, this month, it's April, we're gonna put this stuff in the factory. So that's sort of the ongoing creative process. And it's never gonna end because the more efficient you guys get at this, the more the firm's gonna grow, right? It's gonna attract more clients, make more money, and then you're gonna come up with even more ideas, how to improve and service people differently. So it's an open-ended process. So we're just gonna get used to that, which is part of the fun. Cool. Yeah, it's just something you gotta get used to though, because now ideas are so actionable, you've actually got to contain it. I've got to curb my enthusiasm and actually turn the idea into production so that it actually is used. So I guess that's the way my value is, is I can get stuff out of the workshop and into the factory. But my prediction is that you're gonna keep maxing out your context because as soon as you get those 50,000 tokens out, you're gonna come up with a whole lot more ideas and max it out again. I've got my dog about that. Exactly. Yeah, but if it can go to factory level, because now the meeting brief is nice for a standard thing, but then once a year we might want a bigger review, which might be something more like this. So we're like, okay, this is what that looks like. Exactly. So we have a big menu of things that Claude knows how to do that are structured. So there's an annual review, there's a family meeting, there's an IC meeting, and then you'd have this big long list of all these things. And the experience as the wealth manager is you're just sitting at Claude saying, I've got an IC, imagine planning your day, I've got an IC meeting tomorrow, I've got this annual meeting with this family on Tuesday. Pull a report to me. It's that simple. You're just dictating by voice, right? And then you get the dynamically created documents. That's what it needs to get to, right? No, I mean, that is amazing. Yeah. So that's what we're aiming for is like, so a lot of the time I go like, okay, just hands off the keyboard is my idea, right? So you just voice mode, you just talk, and the AI does all the tricky stuff. As soon as you put your hands on the keyboard, you've kind of lost already. We have to, but it's ideally you just want to be talking and being, and stuff being done for you. That's where we want to be. Cool. So do you use voice mode? So do you do that, yeah? I don't in Claude much because their voice mode isn't very good. The transcription is poor. So I use an app called Super Whisper. So it's basically an add-on to your keyboard and you just push a button on the keyboard and it transcribes separately and dumps it into whatever context you're in. So it throws it into Claude, your email, your WhatsApp, whatever. And the transcription is really good. So I use that all the time, yeah. Okay, okay. And then just closing up here, maybe something to scrap for phase two, which we'll clarify in like week three, week four. But if part of that can be going into like our current Claude setup and then just seeing if it's like best practice or anything, no, just like our connectors, what's custom, what we've got set up, just general like hygiene for Claude specific. Very happy to sit together on that, yeah. So let's get a few things kind of going and structured and then, well, very, like the thing is about the project is it's, like I said, the main thing is that it's an ongoing, creative, generative process, right? So it's never gonna be that you're not doing weird, new stuff in your Claude interface. You always, as soon as your time frees up, you're gonna invent more stuff, right? So very happy to help. But I just wanna get some stuff off your plate and put into the factory first to free you up and then we can kind of look there. So, you know, the real thing to do at this point is just to get me the docs. Let's sign the NDA and the agreements. Get me as dump as much as you can. I'll then go away and work through it and get started. And then, yeah, this is a really exciting project. So looking forward to it. Yeah, we're really looking forward to it. And then also they're bringing in like co-work because that can help so much, but I haven't engaged it yet because I don't know if it's still like in research mode or, yeah, there was some hesitation I had. Yeah, I think the hesitation is sensible. I've got a lot of clients who are using it, some amazing results and some like huge frustrations. So, you know, you do a whole lot of stuff in a thread and suddenly it just, the thread dies and you lose all your work. I've had a few people report that. Or, you know, so it is epic, but it's also early in research mode. And if you're relying on it to get work done, it's a bit scary because you never quite know when it's going to fall over. So I mean, I definitely play around with it. Do you use Cloud Code yourself much? Okay, so that's really interesting, but we can get to that another time. So yeah, perfect. Perfect, and then we'll have a separate discussion on MCPs because we went down that rabbit hole and we came to some astonishing realizations of what it could lead to. Your sub-stack was one of the inputs. And it's, but is there a debate between MCP and just like command line? You know, what do you think MCP will be in that? Look, it's very much in flux, right? It's very much in flux and things are going to change usually. But MCP is like the way to go at the moment for securely connecting your live Cloud chat to your database. So that's the industry standard right now. It's definitely changing fast. It's definitely going to be a lot more that's going to happen, but for now it's a good option. No doubt it'll shift a lot over time, but for now it's solid. It's got lots of problems, but it's sort of the best we've got to work with. But the implications are amazing. Yeah, no, it doesn't matter. It kind of covers both, so, yeah. Okay, perfect. So guys, I'm heading to my fourth, sorry, I've got another call starting right now. I will catch you soon. Graham, thank you so much for your time. Don, thanks for sitting with us. Guys, catch up soon. Ciao. See you soon. Bye-bye. Cool. Okay, so NDA and I'll just share those files with you and then I'll wait for you too. Okay. I'm really looking forward to this. Lovely. I'll get you the NDA immediately and then the agreement shortly thereafter. I think we can start as soon as you've got the NDA. Like, we've got a handshake agreement here. We know what we're doing. We don't need to stand and sign. But I'll leave you with the NDA because that's a compliance requirement and then I'll start working. Correct, yeah. And then we can sign the commercial agreement. So just like, let's cover our bases, but we also don't need to wait once we've got the NDA done. Yes. Okay. And then during the phase, how best do you like to communicate? Is it going to be by email? Do you want to have like a weekly call or what works? So, I mean, I want to kind of hassle you as little as possible is my main objective because you've got a company to run. I mean, funny enough, a lot of my client interactions are over WhatsApp. Like people send me voice notes. So I think that's like the most convenient. A lot of clients literally just voice note me. They come up with an idea. So you're sitting, it's 2.45 on a Thursday afternoon. You've got an idea or a concern. Don't even bother. You don't even have to call me if you don't want to. Just send me a long voice note. Usually like three, four minute voice notes. And it's an easy way for you to get stuff out. And then we can chat on WhatsApp. That's a lot of my clients just like send me quick bounces on WhatsApp. And then there's a phone call if you want to actually like get into a back and forth. And we can also set up, you know, proper sessions like this to demo stuff and walk through stuff. But my, you know, it's just overall the vibe is, I'm happy to give you whatever time you need, but I'm very conscious of people's time. So if it's a call, it could be a WhatsApp, then make it a WhatsApp. If it's a meeting, it could be a call, make it a call. If it's a meeting, it's great to have a meeting for something like this, but unless we need it. So I try to avoid a scheduled routine meetings just because everyone's got like massive Zoom fatigue. But I'm available to you and communicate to me in a way that's most convenient to you is the general vibe. I find lucky people to send voice notes, to be honest. But it's your choice, what you feel comfortable with, yeah. Okay, perfect, because it might be, so, because I really do want to be hands on deck here. So if you do want to, this is important for me. So if there is feedback you need on that, like I am available as well. I'll arrest you, but don't worry. I would prefer that because then I can work through it because I think the time invested here getting this foundation right just unlocks so much. So I'm really keen to, so don't worry about my time. I can manage that. And then I'll try not to, in this case, this is not the idea generation. The ideas are there. It's all been there. So I'll restrict the idea generation for now. Yeah, I will. Hang on to it. I'll just have to go into my cave and figure out how to make it all work. So you just have to be there for me while I do that part. Looking forward to it. I'll pop you that in the air and you can go from there. Cheers, Ed. Thank you. Cheers, everyone. Bye.