Hello everyone, I'm going to be talking about AI -driven event analytics in real -time dashboard. My name is Diogo, I'm from Portugal and I have a background in cybersecurity, but I'm also passionate about AI, that's why you may have seen me at some of these events. And outside of work, you'll find me playing basketball by the beach, so that's a bit about me.

meet.

The Challenge of Turning Feedback into Action

The problem, 40 % of event organizations struggle to turn event data into actionable insights. This was conducted in a study in 2026.

And this image here is conveying just a bunch of feedback forms which people enter after an event, which is already a good sign that people are engaging and want to improve an event but then the organizer needs to look at all of these scattered feedback forms and somehow make use of them for the next event.

They need to compile this and find actionable steps to make their next event better.

A Real-Time AI Feedback Dashboard

So how do we fix this? Well we've built a solution for this as part of our event management platform and I please ask

everyone to scan this QR code and do a favor to Andreas and the organization and give a short feedback about this event it's just a one -minute feedback form don't overthink it there were wonderful talks by Rick Udum and blessed so you can comment on those too and yeah what I'm gonna do in the

meantime is open the dashboard that we have set up so we've got it over here and AI is going to get in action so we've got one person that's already responded so I will say as well that the AI is going to require at least five responses to get a first dashboard so that it has enough information so as you

can see here already four people answer that's great five and as you can see now on screen it's generating the AI feedback dashboard so here we've got our first dashboard and it's going to give us a it's a bit difficult to see now but but the main point is going to be where to focus.

So where to focus section is giving us exactly what needs to be focused for the next event. So here it says maintain educational focus. So it's actually keeping something because I guess the feedback was very good.

We have here 4 .6, preserve technical depth, clarify any ad content to avoid confusion. But yeah, these are just some minor feedback and it's great to have for next event.

How the Dashboard Works

then i'm going to explain how this works so let's go back to the presentation how do we build this

Prompt Engineering and Structured Outputs

end -to -end the first thing starts with prompt engineering so we're passing all of your feedback responses into an ai an ai model and we really wanted to respond as well as possible and the first key is to we thought okay how do we give it all these feedback forms what questions do we ask the AI and the first thing is to make it structured to tell the AI I want to

receive the data exactly this way so sentiment oh it can only answer with positive neutral negative positive points an array of strings are really being specific about the data type we give it and like this the AI will have less of that non determinism that we were talking before of what I mean by

non -determinism is if you ask it a question it may answer differently every time and by having a structured way that it needs to output we're going to try to reduce this as much as possible

Pre-Computing Metrics and Protecting Privacy

also as part of prompt engineering we want to have pre -computing i have this image here and it's to explain more or less how people use ai nowadays the models nowadays are extremely extremely powerful, right? And you just ask anything to AI. People are no longer users in their brains. And that's why I included this image.

But the truth is that AI isn't the best thing for many tasks. And to do a simple two plus two, Python script will do a much better job because it's deterministic. We get always accurate answer, less token usage, and just faster. It's going going to be doing it instantly.

So the idea is use code for pre -computing the metrics. When you guys answered those rating questions or if there were some sort of multiple choice question, we would create all of those metrics, averages, anything that is numeric, we're going to be doing in code and then we pass those numbers to the model.

And to be even more intense to the model, we're going to forbid it from authoring any number. So we're going to tell the model, hey here is all the data we've given to you and now you shouldn't invent any number up. So that's the second prompt engineering technique we used.

And thirdly is linking to the security perspective and it's PII redaction. So personal identifiable information is what it stands for and

And you guys inputted a feedback form, and those forms were anonymous, so your data is not going to the model. But, for example, in one of those open answers, you may have included an email or a link, something that we also don't want to enter the model.

So what we do is we use a regex, which is a technique in code to essentially match for certain patterns. So if you have a huge block of text, you can match for emails, links, phone numbers, and we remove those before passing it into the actual model.

Model Selection and Provider Resilience

And on to the fun part is the model itself. So now we've done everything, input, boom, we send it to the model and we're using a tool called Open Router. And as the name suggests, you may have heard of this tool, it's a router, so it allows you to try out different AI models through different providers as well.

And this is a very key part because many times you want to have that availability that if one provider fails, let's say Anthropic, you can fail over to another provider like Amazon Bedrock. So having this possibility is key.

But to be honest, the main advantage for us was to to be able to have a lot of models to test this feature with, which models create a better dashboard essentially, or which ones have a faster inference. So this was a great tool to use, and it's extremely simple to set it up, even if you're not that technical.

It's really just this block of code. the first line here is just calling the API endpoint so sending the request to that URL and the only thing you need to change from here is this open router API key which as soon as you sign up to open router it gives you that API key you just copy paste it in there and then it's as simple as switching here the model which you want to use the parameters here might vary a little bit but aside from that it's it's really simple okay so now we've passed all of

Validating AI Outputs for the Dashboard

these feedback information into the model great so we're done right you know we passed it it gives us an output but actually there's more to do because to display it in the the dashboard we which I showed there has to be processing

first one being basic error checks so let's say oh maybe the battery ran out there yeah yeah I'll speak okay so so the first is error checking so there's going to be basic error checks that we want to do this is as simple as I mentioned before if a provider ran out if a provider is out or if we reach the end of that context window of the model and for some reason the okay and for some reason the output gets truncated, gets cut. So if that happens, we want to have those basically just sanity checks. But we want to go further with this.

The second one is going to be parsing the JSON. The structured output which I talked about at the beginning is what the model is going to return. But this is not yet ready to be displayed in the UI we need to first convert those JSON attributes into actual variables which the front -end code will be able to consume and we want to be even more

specific because we mentioned the issue with determinism and AI potentially giving us a different output to the one we desire and to mitigate that we're going to check that the exact data types we requested are going to be the ones outputted. So as I said before, sentiments, those three positive, negative, neutral, is that exactly what came out? Or for example, if we have the other field, which was a string, is that what really came out?

And there's a library, TypeScript library called Zod, which you may have heard of, which is great at creating these schemas, which you can then compared to the output of the model.

Schema Validation and Hallucination Checks

So that's the third check and finally we want to take it one final step further which is has the model done some sort of hallucination.

So in the analysis maybe you didn't see so clearly but I'll actually show it quickly we can see that at the bottom here let's go down we We actually got 11 responses. Nice.

And as you can see, it generated a new updated report from 10 responses. So it updates every five responses over here. So we've got a few more pieces of data.

But at the bottom, we've actually got some quotes that you guys gave and you can confirm that you wrote. And we want to make sure that these are 100 % real quotes and that the AI didn't make them up. So what we're going to do is in this final step, we're going to compare the quotes that the AI chose and output with the ones that were initially passed in. So we do that comparison on the server side. So, yeah, that's the end to end flow and how we build it.

Using AI Where It Adds the Most Value

To many of you, it might sound really simple, but I think this is one of the keys, not only for startups, but for larger companies. companies and it is to really know when AI is the right thing to use and when you can do for example a pre -computing step because on a large scale the amount of tokens that you're going to be saving if you know when to implement each tool is key.

For example if we're doing maths as I said deterministic check easy you can do that with a python script but if you want to have a reasoning reasoning task, that's where AI is needed.

So yeah, that was my presentation. Thank you.