Introduction

I started out this journey and I wanted to build something small and part of what we do is we do some coaching work. with other CTOs where we help them to understand, you know, oh, you could do this differently or you could do a little bit better.

And so as part of that coaching work, I do calls and we record the calls in Gemini and we get an auto transcription. But that kind of saves all of those auto transcriptions into like a Google Drive folder.

And it's a little bit of a nightmare for me to then go in and find the right call from the last week and all that kind of stuff. So what I thought was, well, let me build something that will take away all of the pain and effort that's involved in that.

Building a Coaching Solution

So that's what I built in our coaching.

And so what this does, again, I didn't write any of this. I didn't write a line of this code. I know what's going on because I'm technical, but I didn't write it.

But what I did was I built this page, which has different coaching engagements on it. And then it has an integration with Google Drive.

And so it goes and looks in my Google Drive folder for where the Meet folder is, where all the recordings are. 1And then what it does is it allows me to assign these notes, which are notes from a call earlier today, to a particular person. And then it makes a call to Google Drive, moves those notes from where it is into the right folder automatically, and it just works.

And I didn't write any code to do that, which I think is pretty cool.

Features of the Platform

You can go in here, you get all of the notes, and then you get a nice summarize button which will call Claude, I think Claude now, and it will summarize all the notes of what I did with that person the last time I talked to them. I think that's pretty cool. So that's the first thing that I did.

Yes, Lou has a question.

Yeah, so it's a good question and so Sometimes it's a little bit trial and error, but Claude is much better at Some things then Gemini is then open AI is and so I think I just prefer the way Claude does note summarization for me. So basically, I would just try it with each one of them and do a little bit of experimentation.

But Claude, I found, is generally better at writing, like prose and stuff like that. So that's why I kind of tend to use it there.

Any other questions about that? Probably not. It's very, very small.

Expanding the Solution

So with that kind of experience, what I decided was that I wanted to build something bigger.

And what we do is we do these due diligence pieces of work where we go in and we look at a company and we assess them on a bunch of different metrics. And so what I did was I built this as a platform that allows our CTOs to go through and answer a bunch of different questions and then do different things based on those questions. Basically, we have a set of questions here that you can add comments to and thoughts to and so on.

And then you can click this button and you get this really cool pop-up. Again, I think all this is cool.

You get this cool pop-up that allows you to take all of the questions that have been answered and send it to any of the models and any of the types of models. And then it gets you a big response back of here's what we think about it. which I think is nice.

You can also, take all of the files that you get, like the transcriptions and so on. You can analyze them automatically and say, OK, well, give me the answer for all of these different questions. And I will go and run the analysis and come back and tell you all of this.

Technical Implementation

I didn't write any of this code. That's really, really important.

Hello.

So that's really cool in and of itself. And then, I'm not gonna wait for that to come back.

And then what we do is, there's like a bunch of other different features in here. And then we go in here and we can generate the report.

And this, is all auto-generated. It's like 50-page report based off of all of the things that we have in the audit.

And I didn't write any code to do this either. And that's cool. I think it's cool.

The Role of Coding Experience

So does anyone have any questions? I know this is like super high level, but I'm going to try and do some code in a minute.

And we're going to see. Yes.

So you were stressing how important it is that you didn't write any code, and I understand a little bit why, but I also wonder, you did your job for eight years without writing any code, you just said, so how much does it differ from how you usually have directed a team to write your code? So the question, I think, is how does what I'm doing here differ to directing a team, essentially?

And it doesn't, really, I don't think. The reason that I think that I work well with this is because I have a lot of experience in directing a team and helping them to understand what we're doing and how to put the pieces together and understand when they're going down the wrong road and stuff like that.

And so I don't think it's fundamentally different to running a team.

Because, yeah. Yeah, OK, fair.

Experience and Adaptation

And that's an interesting point in itself because what I've experienced is, and we do like a bunch of training for engineers now on how to use AI tools, and engineers seem to struggle with using AI tools because they're not CTOs, right? I know a bunch of CTOs who do really well with it because they understand the abstraction and you have to give away a little bit of control and a little bit of like, yeah, just a bit of control in general.

And so I think that it's not really that different from operating a team.

There was someone at the back first, sorry folks. So to make the report itself? No. That's a very interesting question. Yeah.

So there's a couple of different things going on. And so the question is, are we using an LLM to generate the report? And I think that's a really interesting question.

So no, we're not using an LLM to generate the report. What we're doing here is we are using a bunch of LLM tools to generate the content in the report.

Report Generation Process

but in a very manual way so it's very, we have a CTO who is putting in their opinions and their thoughts into the system and then it is taking all of that and putting a custom prompt around it, sending that off to an LLM and getting some content back and then the CTO will check that content and see does it match what they expect it to be, like changes the opinions and so on, is it hallucinating anything and so on And then we save all of that content and then we take that content directly and put it into the report.

But what I think is particularly interesting is the report is actually generated with what I would say is a bit of an esoteric PDF generation library that is basically designed for academics and that I didn't even know existed until I started looking for it.

And so what I then did was I took all of the documentation, well actually I took the one URL for the documentation for the library, I gave it to Cursor, and Cursor is a code editor tool that I'll show you in a second. I gave it to Cursor. Cursor read all the documentation, and then I was able to interact with Cursor and say, okay, I want to do this thing with this library.

How do we do that? And it said, oh, yeah, okay, well, this is how the templates work, and this is what the language looks like, and okay, I'll write that bit for you, and I'll write that bit for you. And I built an entire template for building these reports without learning the language that we're actually building the report in.

I think that that's just incredibly powerful and really, really interesting because I didn't even know the thing existed before, and now I'm using it. But I couldn't tell you how to write in this language at all.

It is a text-based language. It's cross-compiled. It does a whole pile of random, random stuff. And I don't know how to do it, but it's really, really interesting.

Yes? No, no, I'm sorry, behind you. He had his hand up before.

Coding Knowledge and AI Tools

You may have answered this in the last answer, I just didn't catch it, but it's very impressive, especially that you haven't written code for eight years, but to what extent do you think your knowledge of code was a prerequisite? But to what extent do you think you looked at this code and sort of bettered it first and I guess my real question is can a normal person who doesn't have a background in coding like me for example how easy do you think it is to do something like this based on just only on a concept or an idea of what you want to do and what you want to achieve

For example, I do a lot of writing in my job, normal writing. And I do cut a lot of cordons these days because I know what is good writing and I know when the NLM is early on. And I always say to clients, you don't want to give this to an intern who doesn't know what good writing looks like because they'll just churn out rubbish.

Yeah, I think it's a really good question. So for those that didn't hear in the back, the question is, how important is it that I know code and know what it's doing to be able to manage it?

for something this complicated, very important. Okay. So if you're a developer, then it's very important that you understand what generally speaking what's going on and you can kind of keep it on track.

There are many times whenever I could watch it go complete, go off track and I know, OK, well, if you keep going down that road, then we're going to have a problem. Let me pull you back from that. Let me check your assumption and so on.

And I think it's the same for the use of any of these tools right now, like code tools or not code tools or whatever. You have to keep a bit of a leash on, if you will. I think that right now you get like a 30% to 50% uplift in your productivity.

But the people that think that it just happens, that it all just works, they're the ones who really struggle with all of this. If you try to keep too much of a leash, you have a problem. If you try to just assume that it's going to do everything, you have a problem. And so for something this complicated, knowing how to code is very important, in my opinion.

Now, there's a secondary question to what you're saying, which is how possible is it for someone like you to write code? Not to do this, but yes, very, very possible.

Using AI Tools for Non-Technical Tasks

So there's a slight tangent now. But have you heard of Replit?

So Replit is like an AI code builder. If you're non-technical, I would encourage you to go and have a go on.

But I can show you something that I built in this, two things that I built in this that... basically I kind of one-shot prompted them and they just came out and it just works and I didn't need to know any code to do this right and so first of all like about well three days ago I I was walking my dog and listening to a podcast and I was like oh the insights in this podcast are really interesting I would really love to throw this into an LLM and so I went home I wrote a prompt to replit and And I said, build me an app that will take a URL for a podcast.

It will download that. It will then send it to a transcription service in OpenAI, take the transcription, and then take all the interesting information out of it. And so then it built this.

That's the database, so that's not useful. But let's wait for it to start. But basically, it built it in one go, or in a couple of goes.

This wasn't part of the original demo. On the plus side, we're going to have not enough time to see all the code, which I'm sure everyone is very pleased about. But if you want to see code later, then I can show you how a code editor works with AI.

Practical Applications

So this takes a podcast URL, and actually, it takes the URL of an MP3, or I got it to take the very specific URL from my podcast app, and it will then look at the web page, figure out where the MP3 is, download that MP3, send it up to OpenAI, transcribe it, and then take all of that and get all the interesting bits of information out. I didn't look at the code for this. I haven't seen the code for this. I could if I wanted to.

And then another thing that I built is I'm back to the UK pretty regularly. I live here. But I built a tax tracker. And actually, I think this is just live. Let's see if this works.

yeah so i built this tax tracker i get this was even more one shot like literally i just asked it to build it for me um and so now this is this tracks whenever i'm in the uk so i don't go over the tax threshold and it also has portugal and you can add other countries and so on um so you can absolutely as a non-technical person build technical things um which is really fun sorry how far do you think we are from doing the first thing

Oh, like a year, two years, something like that. Yeah, maybe a little bit longer than that. Yeah.

You had a question. Sorry, you had a question first.

Challenges and Considerations in AI-Driven Development

My question is more related to the making ability, the long-term viability of a product that was used entirely with an AI. Because one problem that I see nowadays when I'm using it on my job is that sometimes, for example, a cursor will spit out 400 lines of code at once, changing. several files and if I go and open up PR with all of those changes I would definitely get some not very friendly replies into my system and I find that sometimes the tooling is so effective that sometimes it makes it hard to follow what it did and go back in versions and fix bugs, etc.

So, I wonder, in your experience, if you built a complete product using MLMs, and how did you go about this, if at all? Did you just, you know, Yeah, I let it run wild because I was building an internal platform for us and I was fine with it, and I don't mind the extra code, and I don't have to interact with other people.

I mean, there's a whole other thing that we can go into, and let's talk about it later. I think the paradigm shift on what engineers and engineering teams will have to do in the future is going to be massive, and we can talk about it more later, because it's not like it's going to be before, but equally, the pace of change is that's the direction it's going, and it may not be perfect, but a lot of tech stuff isn't perfect, and that's the way it is.

Viability and Engineering Paradigm Shift

I'm going to take one more question and then we're going to have to wrap up. Yes?

For what I built? Well, anything that I would have used would have been open source libraries and stuff like that.

As for the intellectual copyright in the code, I'm not best place to, like Lou is going to be the best place to answer that question. Sorry, I know that's not what your talk is about, but I'm not best place to answer that, to be honest.

Yeah, so I would just be opining very badly on something that I don't really want to do. Sorry, we're going to move on.

Come and ask me afterwards. We're going to move on.