So I'm Quentin. I work in one of those companies who work on all the cloud and AI stuff, but I just want to say this. Anything that I share today is only my opinions and not my company's opinions. That's just for the disclaimer.
What I'm going to present today, and I'm super glad that both Bruno's went before me because the first Bruno introduced what an agent is, and I'm going to go into showcasing you what you can do with Hermes, which is one of the agents that you mentioned.
And then about the curiosity, because that's where it started from, how to build something for me, not for like a big use case, but just for me and my partner and our household to manage our money. And I'm going to talk about money.
So for cybersecurity, security, we're gonna have a lot of fun. So well, I'll do it old school then. Okay.
Before I start, so I used to be a financial advisor when I was in Ireland, this is not my job. But this is a side hustle kind of thing.
And now that I came to Switzerland, I kind of have to start from scratch on the financial knowledge because everything is very different in Switzerland when it comes to personal finance, but I'm very passionate about personal finance. So how do I bridge the two with AI and personal finance?
And so everything that you see is dummy data. That was the hardest part for me in this presentation, is making the data kind of shareable, because otherwise that would be my financial data with my partner's financial data.
But pretty much every month I have a routine of reviewing my finances. Every quarter I have another routine, again for finances but different purposes, and then every year. Again, I'm a bit OCD on finances because also financial advisor.
But the idea is that every month, going through all the statements, where is my money? Is it invested in the right pocket? Is it optimized from a tax perspective? I know Switzerland is quite a good place for that.
And then also, what should we do and how should we be more free when it comes to how I use my money? meaning like I can forget about investing in everything because I know that it's under control. And so what I have left, I can do what I think is relevant for me.
And so this is pretty much what took me a lot of time before every month, one evening, a bit more to just go through all the statement, putting that from different places, not having MCPs for all the banks, so not connection for all the banks. And my approach was to try to find a way through Hermes agent to centralize that and have that doing it for me in like 15 minutes or something like that.
Okay.
So that's kind of the outcome that you get out of it. Now I know it's a bit overwhelming if you're not into finance, but it pretty much gives me like, this is what is needed from me this month because either the information is stale, so it's missing this information or there's some actions to be done.
So for example, pillar two, it's an important one, is the certificate of today. day then it gives me also based on my investment strategy I want to be financially independent for those who don't know it's the fire movement so how do you be financially independent as soon as possible to then be free to not work if needed that gives me kind of an idea how far I am again all of this is dummy data then you have a different pocket so much I have in cash etc etc I don't want to spend too much time on the finance side of things but that gives you kind of where this is coming from.
Where I spend my money? Is it spent in the right pocket? And then it can tell me inside, like, hey, it seems that you're spending more on this. Is that what you're looking for?
And then the saving rate, which for me, and this is not financial advice by any stretch, but something that everyone should know, what's your saving rate per month or per year? So how much do you put on the side? And then you have the different split of portfolio.
So, why didn't I do that with any other solutions like ChatGPT, Cloud, Mistral, others? First was privacy for my documents.
So, I'm going to spend, because I understand there is some question there, so I might spend a bit of time to acknowledge the privacy leaks that I have right now. But privacy, I don't want my bank statement or all my information to be in the chat. And what I mean by being in the chat is if it's put in the chat, it means it's hosted it behind in the cloud of ChatGPT, so OpenAI or Entropic, which can be AWS, Azure and others, which I might not want to have that.
The memory. So I don't know about you, but I feel that in the chatbot, it's getting better, but it loses a bit between the conversation. So it doesn't really know me and it doesn't get to know me better, which means I sometimes have to repeat stuff. So I'll speak a bit about that.
Consistency, this is still an issue, but it's at least better for me with my setup. You ask the same question, but the output is different all the time. If there's a slight difference, that's okay because it's not deterministic, but still, there's methods, et cetera, that are different.
And ease, where it has to leave where I am, so that was my phone or my computer. So, I really needed an assistant, not a chatbot, because for me, I need something that's running every day.
So, where it lives? This is me on my phone. I have Telegram, which is the gateway. So, that's the connection I have with my setup.
And at home, I have my mini PC with RMS running. So, the mini PC is on Linux, which is slightly better from a security perspective, in my opinion.
Union. And then I have a VPN. I don't want to be too technical, but a VPN that is not exposed to the internet that allows me to directly query on my mini PC for all the workloads I have.
So that's my mini PC. There's not much else than that. And you can see I also like to game a bit.
So how that works. So again, in the middle, we have the same setup up, as I mentioned earlier. So language models, we talked about those.
So I use my GitHub Copilot license, which gives me access to all the different models. I want to call that out because from a price perspective, the cost will be vastly different between the models that you use.
So there are best practices. I can share the best practices when it comes comes to using models. To plan, use the most, let's say, the highest, most frontier models. To execute, use the lowest. This will balance the cost.
I use my GitHub Copilot license. For memory, I use what's called HintSight. This is an open source solution. What's good about it is that when I chat with my mini PC, of course, it memorizes facts, but then it runs an embedded model, which means the model within Insight that just runs throughout the day, checks the facts, and makes sure that it makes what is called observations.
It makes the links between the facts and previous facts that it knows, and then it creates a map. I cannot show it today but there's pretty much a very cool mapping like a brain and it creates all links between all the information and the good thing is that at the beginning it's a bit dumb like you know but it gets way better very fast the more you use it because it creates all those connections
and so there is my bank of knowledge my partner's bank of knowledge which i cannot see but she can use it and there's a shared bank of knowledge which we use both so we can both query nothing gets written without my approval that's my rule for everything and then I have skills
so skills are repeating procedures or repeating prompt the different tools that Bruno talked about I have to do is do but I have also my google drive so now cyber security is going to be crazy I have my google drive I have a ibkr mcp so that's the trading platform I have everything Everything I can connect, I do connect, I have Google Sheet and everything.
And then I have scheduled shops, so every morning I have a brief of my day, what should I think about, what are some ideas, et cetera, et cetera. And then I have my monthly review, so the one I was mentioning about when it comes to finance.
Is it clear so far? I hope I didn't lose too.
One thing to know, so what is important? Because we're talking about finance, the setup is important.
Usually, what I started with is like saying pretty much the prompt is run on the highest most frontier model and have a look at the tax treatment in Canton X, the Canton where you live, and create a list of the top things that I need to know from a personal finance perspective. Then it goes through it.
My goal is to create an investment policy statement. If you don't know what that is, it's pretty much like putting on paper what is your strategy energy when it comes to finance. Highly recommend that just on a personal level, you don't have to do it with any AI, but it helps you stay the course even if the market is changing.
Why am I saying that? Because it's what the AI is basing every, let's say, recommendation or I don't want to say advice because it's not advice, but like steering, it's based on the investment policy statement that I created. It is one file that the AI is always going towards to see, okay, is that going into the direction that you set? It covers allocation, risk, how much cash, et cetera. That's the point I made. It applies the policy. It doesn't improvise advice because otherwise, it can go crazy and be unrelated.
The good thing though is that because I have the insight memory, the more I use it, the more it understand me and the more it gets, let's say, tailored on the advice, but I don't need that anyway, because I need an investment policy statement. This is only because we talk about finance.
For all my other stuff that I'm not going to talk about, only my memory is enough, because now it knows me enough, how I work, what are my preferences. I already kind of touched on that.
I want to demo. Okay. So I'm going to demo how it looks like.
So I prepared a static demo in case it didn't work. But I think it works, so we'll see. So this is how it looks.
I called it LifeOS and this is a dummy account I used for it because there are other buckets otherwise. So I can sign out, sign in, like any kind of web app.
This is hosted on my mini PC, so this is not hosted online. I'm just, as I mentioned, I'm using Tailscale which is a VPN straight to my mini PC right now.
From a cybersecurity perspective, the only protection is password which is hosted on my mini PC but it's local and so everything the example is Alex there so this is what the page for Alex look like so you already saw like the screenshot but this is what it looked like right
cash how much investment pension again all the data are just dummy data but this is what you you can get very easily it took me like literally 30 minutes to to get that up because I had the IPS already in the system, et cetera.
And then it gives you a visibility on like, what are your expenses, et cetera, et cetera. One thing also I worked on is, cause the initial thought was like,
how do I work with my partner on like making it easier to know like, what do we spend money on? How do we split things, et cetera. So you also have a mechanism to have the shared. So everyone has their own space, and then there's a shared space.
And so this is kind of the asshole visualization, where you can see, like, okay, what are the things that are expensed in tax foncière in Annecy? I don't know how it took that. How is it split? What is the split of the costs, which can be important to understand?
And also what's the evolution of the property value, which is can query online based on different database depending on the country etc etc okay.
And then there's the monthly review so unfortunately the monthly review is difficult for me to demo as it works because I would need like statement and things like that which I cannot create
like dummy statement for everything but pretty much the idea is like you upload the files, it analyzes the files.
I'm going to go back to the privacy because that's an important one.
You upload the files, it analyzes the files, and then it tells you like, okay, this is where the money went because it knows the categories, it's an AI, so it knows that Migros is grocery, unless it's Migros Bank and then it's something else, but it knows all of that, so you don't need to hard code this.
Now, I would say before anything, from a privacy perspective, there is an issue. Maybe I'll go back to my slides because I think it's important for everyone's benefit.
Everything is local. This is local. Everything is happening on the mini PC there, that part.
What is not is that part. This is going somewhere else. What we call the inference, so the intelligence, is going somewhere else. That's the leak of privacy.
You need to be comfortable with that. That's a risk assessment I did for myself, and I was like, you know what, considering the benefits, I'm okay with whoever is the model creator, like Anthropic, OpenAI, Microsoft, Google, whoever is over there, I'm okay if they see some of those data.
You might not be. That's on you, and you might not be. But I want to acknowledge that this is the point of vulnerability I have in my system.
Now, that's how it is today because I want have access to the Frontier model without paying for big hardware. Now, you can do everything locally. There's no thing against putting a model there, just like I did for Insight, meaning the model is running there and all the inference is done there.
To do that and to have the right level of intelligence, it requires a way better mini PC than the small box that you've seen, and I don't have the money for it. That's the risk assessment I I did for myself.
But I want to highlight that because when you put things in chatbots agents, most of the time, unless you have a local model, which I think 95 % of people don't, it goes somewhere. So I just wanted to highlight that for everyone to be very aware of what you're doing with AI.
That makes sense. Yeah, that's it.
We can have a big man full of time pizza is coming.
What I would say is like, I picked one friction friction point, which for me was the finance, and it's also an area I know about, so I can kind of challenge what the AI is telling me.
Just like Bruno said, do the same with your area where you know a bit, try to challenge a bit the AI, but build something that makes it easier for you. So I would say pick one friction point, build one routine for it that works for you.
I'm using Hermes, why? Because what I said earlier, and also for fun, because it's good to tinker, know a bit better, know what's behind the scene. It's also a good way to have more control, as we discussed.
And I like also the idea to not provide my credential out there to other, let's say, providers and have it on my box. I didn't touch on the credential sharing with LLM, but that's for another day. We can discuss on security.
And yeah, that's it for me. I hope that was clear.