Introduction

Thank you very much for having me here this evening.

As you can see, we, me and my partner who isn't here this evening, are wealth engineers.

This is kind of, I suppose, a little bit high level. So if there's more technical people that want to dive into some of the technical questions, please, by all means, ask me as we kind of go along and we can talk about it.

From Startups to Wealth Engineering

But essentially, I've been in startups for a very long time. I've been through kind of Web1 era, you know, flogging products from Eastern Europe into the UK, through into crypto and Web3, and now working significantly in AI products and combining different bits of my life together.

And that's actually really where Sentry started from and why we ended up doing this and bringing the human element back into AI, as we like to put it.

A Personal Problem With Advice

So I moved here about three years ago. When I moved here, I was given horrific

tax advice from my uh man in the uk i was given even worse advice from my man in the uh or woman actually in the uh in in portugal and uh and then wherever i had yeah a few assets and a bit of other advice They all kind of laid it upon me and then did nothing with that particular advice.

It was all left up to me to work out really how to pick through, how to structure my life, my wealth, everything that I'd kind of built up in my path to getting here.

Why Generic AI Fell Short

And I turned to ChatGPT, like everybody would in that kind of scenario. And then I got even worse advice from ChatGPT because ChatGPT is prone to hallucinations and all sorts of other issues.

The Missing Middle: Not Ultra-Wealthy, Not DIY

Now, if you're super uber wealthy, you can solve this issue. You can just hire a whole team of people who work specifically for keeping you wealthy.

If you are on the other end of that scale, like you've made a few quid over the years, you've done okay, you've made a few investments, you've exited some startups, whatever that is, you're kind of left up to your own devices. And you can go out there, you can grab these bits of advice, but it's very difficult to then work out what to do with that.

And so that was the issue that we were kind of faced with.

Founding Sentry and Defining the Problem

And I met my co-founder about a year ago. He found the exact same problems.

Both, yes, live in the leafy, lovely suburb of Cascais. Yeah, we are those types of people who come here to Portugal and live in our little ivory towers.

You can see the eyes and rolls over there.

Sorry.

Beyond Cascais: A Global Need

But it is actually a bigger issue beyond just the Cascais population. It's actually a fairly large issue to about 75 million people around the world, which is why we wanted to tackle it.

Our Solution: Human-in-the-Loop AI for Wealth

So our approach to this was to really build our own bespoke AI system, built on a number of different frameworks and models underneath.

Right Model for the Right Task

And the reason why we ended up doing it that way was because we wanted to 1Utilize the best model for the context, which if anybody's done any more kind of playing around with AI is a fairly major issue.

If you've ever tried to build a, probably not many people in the crowd have done it. Has anybody tried to build a cap table with AI? No, there we go.

I can tell you from experience, it does a really bad job. Cap tables are the breakdown for a startup and which shares owned by which people. If you build that in AI, the simplest of mathematics, it can often get wrong.

And that's because it's a large language model. It's not actually built specifically for algorithmic issues just like addition and subtraction.

Earning Trust in Sensitive Domains

But we built our own AI system, and that would be fine. But it doesn't really provide a lot of trust.

When you're dealing with these kind of wealth, it's actually primarily a big family issue. It's a big personal issue. A lot of people don't necessarily like talking about it.

People definitely don't like putting it into open AI and other systems that are quite broad and they don't know where their data is going to. We took the approach of taking a human and combining it with AI. So why is that important?

Building the Financial Digital Twin

So context is where we start from when we're looking at building up these pictures of people and how our system works. And essentially what we're... looking to do when we're building out a context of somebody is really understand exactly where their tax residence is, where their physical location is, their citizenship, what type of assets they have, what type of relationships they have.

Have they been divorced? Have they been remarried several times? All of this builds up into a massive web and context on an individual.

And that's very important for us because it means we build this kind of like financial twin or digital twin of their financial life.

Surfacing Risks and Blind Spots

So to give you an example, when we build up this map or this web of a human, we start to see very... clear issues like this property over here might not be written down correctly in your will, therefore your wife may not be able to receive it in the event of a death, or there may be like a massive tax liability if you were to die very suddenly. And so we start to see these kind of patterns and issues that highlight out to us.

From Graph to Opportunities

We then take this graph of a human and we feed that into our systems that we put together. We start to find these various different opportunities and efficiencies.

Why We Call It Wealth Engineering

And that's really why we call this wealth engineering, as opposed to, this is not wealth management, this isn't tax advice. What we're trying to do here is to get a really strong underpinning of how a person's estate and how somebody's wealth is actually built upon. Because if you don't have a really strong underpinning, then any wealth you build in the future, you're just going to erode.

And because there'll be all these issues that come up where you move to a new place and suddenly there's a new tax on top of you and it's not been structured in the right way and suddenly where you thought you were going to be making 5% a year on your investments, it's actually minus 12. And this has happened, this has come up in our issues.

Planning Around Life Goals

Not only when we start to kind of build out these opportunities and efficiencies with inside the graph, we then look to kind of combine that with what the individual wants to achieve. And so we then use multiple different agents in the background in context to then build out a plan based on what it is we want to achieve.

So maybe they want to retire in 10 years. Maybe they want to make a nest egg or pay for their kid's college education or whatever it is.

Mitigating Hallucinations and Protecting Data

Now, very much like you were talking about your ways of looking for different AI tools out there, we have a very specific way of how we protect the data that we're giving to people.

A Daily-Updated, Jurisdictional Knowledge Base

We mentioned about the hallucinations earlier. One of the ways in which we mitigate hallucinations is we have a very bespoke database for us that's updated every single day on

tax information from every single jurisdiction from around the world as well as news information about what's happening and Contextually we can kind of forecast as to what? Might be happening in the world that we can then update our customers based on that information that database is kind of just really the pure knowledge, that is the core of our system and allows us to be able to do what we want to do in terms of planning and engineering for the life events.

From Weeks to Minutes

To give you an understanding of what this leap is for most people, a normal tax lawyer, which is the traditional way of doing this, would take around three weeks or 15 business days to do this process.

They would gather the information from a user. They would then look for inefficiencies and opportunities. then re-present it and put it back to a person.

And even if they're using AI in all of that, it still takes them an inordinately long amount of time. We do that entire process in about 15 minutes.

Cost, Efficiency, and Model Choices

So I mentioned about using multiple different models and multiple different providers. One of the reasons why you wouldn't want to use like a major large language model for doing this is just because it's really inefficient and it's really cost inefficient as well.

So if you're doing simple additions, There is no reason for using LLM tokens and running APIs in the background because the costs build up.

Even when we are doing something like this, we were actually just doing the maths this afternoon just to work it out. It costs us roughly about $25 to $30 per plan in just pure tokens alone. And we're using a multitude of different LLMs.

predominantly we use uh uh chat tpt5 for um when we're doing yeah doing that the large kind of reasoning and when we're interrogating and working with our clients to understand their their their context and their positioning and then we'll use uh you know older models and um you know more simpler uh yeah kind of uh claude etc to um produce some of the documentation that we then give back to the client.

Keeping Humans in the Loop

So talking about hallucinations and how do we mitigate for that, that's really where our humans do come back into the loop. So we start with our main database of information and we operate from there.

We then have our welfare engineers that come into it, and then they look to put together a final solution that has been tested and checked before going back to the customer.

Now, one of the processes that they do during that is always any information that they give back to the customer provides more context to our own LLMs in order to... give a better result for the next customer.

So every single sequential customer gets a better possible result from the information that we've learned from the previous customer.

Case Study: Retiring in Portugal

Right, let me give you an example of how this works. So we had one of our customers come in recently. They decided to retire here.

They've done extremely well, far better than I've done. And they had been here for about two years. They put their head in the sand in terms of trying to think about what to do about tax and how they should be structuring everything coming in.

At the point when we were able to talk to them, there was many things that had happened that we actually can reverse. However, by sitting down, processing their estate, working out how everything connects together, understanding in context what they wanted to achieve in their retirement whilst they were here, we were able to save about 25% of their estate.

Now, without going into numbers, this was like a pretty large number. um and uh yeah they've been able to you know now pass that fairly large number hopefully onto their future loved ones uh or at least have a very nice retirement uh for the next few years

What We Learned Building This

I wanted to go over some of the learnings that we've made. One of the things that we thought we had when we're coming into this, we've been building this for about a year and a half. We thought we would be able to knock this system out in about three months.

Turns out it's actually really difficult to build AI in context systems.

Model Churn and Complexity

It is getting significantly faster, but it's the understanding what the benefits of each model are and then trying to keep up with all of the models as they come out and they change every single week, every month. That has been a huge drawback for us.

The Human Factor

And then, of course, the actual human element.

We're running businesses, right? We've still got to think about humans.

Now, we run an extraordinarily lean ship and that's one of the benefits of AI is that we're able to. No doubt if you didn't have AI, you'd be at least 22 people by now.

Hiring AI-Native Experts

We, in a similar vein, are able to cut down on the amount of people that we have in-house operating on this, but also able to any individual that we bring on board has to be AI native.

And that is our massive drawback, is that we cannot find tax lawyers, for example, that are truly AI native and recognize the benefits uh of this is a huge you know education barrier for us so that we then have to get over when we're bringing people into the business so the ironic uh thing about our company is that yes we both add humans into the loop but by adding the human back into the loop we also create a rod for our own back

Why Humans Still Matter

And the major reason why we do that and have the human is our differentiator, is to build that trust and is to make our AI a little bit more humanoid.

Conclusion

And that's it. That was all I wanted to say about that.