Best AI Hackathon Project Ever | Unicorn Mafia Wins $100,000+

Demetrios Brinkmann [00:00:00]: All right, you all won the hackathon. Congratulations. What did you build?

Marissa Liu [00:00:04]: We have built a solution to this pain point that we've all felt, which is booking group trips. That euro trip summer. Euro summer that you're looking for with 10 of your friends is actually very difficult to coordinate if anything becomes its own full time job between all the conflicting preferences your friends have to the flights that you need to book, then the hotels that need to follow afterwards. And so we built basically an AI travel agent which coordinates among your group of 10 friends everything from the ideal destination, takes in all the preferences from your friends and comes up with the best travel plan. But not only that, it also does the last mile, which is booking all of the flights and the hotels and even ordering food, checking the weather. So, yeah, we whipped it up for the hackathon and shipped it and we're excited to see like 19 or so people sign up for the app. We have revenue. We don't think, like vacations should be a project.

Marissa Liu [00:01:10]: They should be a relaxing kind of activity, not a planning project.

Demetrios Brinkmann [00:01:14]: What were some of the technical challenges you ran into as you were building it?

Ana Shevchenko [00:01:20]: Of course. Yeah. So we actually iterated quite a lot on our design. Like the very first challenge was, okay, what is the system going to look like and how much can we actually build in these, like, five, six days? Right. So we ended up changing quite a lot and we came up with this pretty crazy big infrastructure. I think one of the biggest challenges was probably like integrating all of the agents together and getting them to use so many different tools. It can get a little bit unreal, you know, it can get a little bit messy. You know, if you, if you have an agent and you let it do anything at all, right.

Ana Shevchenko [00:02:07]: You have to coordinate it. Like, you have to put a lot of guardrails and I think we did an amazing job at that.

Charlie Cheesman [00:02:14]: Yeah, I think it was, it was pretty cool. I mean, for a lot of this, we were literally at the tech frontier in terms of what's possible with these tools. So, um, Marisa was doing payments and getting an agent to sort of securely pay for on, on your behalf with no human in the loop is actually quite difficult. So we managed to do this with this cool tool called Skyburn and then passing in credit card and all of this kind of stuff. We also looked at stripe and there are various APIs that are just coming out where you can do this. That was super fun. Orchestrating the agents as well. So.

Charlie Cheesman [00:02:44]: So being able to interface with it multimodally so you can Call it, sort of talk to it via WhatsApp. You can have it on the phone call. You've got an orchestrator that's sort of doing multiple tasks and then calling all of these other agents that are then trying to chat to you. That was pretty fun. Another thing which is they've just released actually ElevenLabs, which is our sort of voice API, just released dial tones. So you can now programmatically interface with like, you know, when you call the doctor and it says dial one to do. To do this or dial two or you need to get through to the hotel's reception and it's, you know, press three to do X. We now have that integrated and that was only released last week, so.

Charlie Cheesman [00:03:22]: So we said, hey, like we need this feature. They happen to be building it. Like they released it and now we're using it as of a week later, so. Very cool.

Demetrios Brinkmann [00:03:30]: Explain more how I interface with this agent. Is it just I drop it in to my WhatsApp chat with all my friends?

Charlie Cheesman [00:03:38]: Sort of, yeah. So. So initially, at the moment we've set it up that you, you will have a little phone call with it and basically give it your preferences and you see it has your phone number and then you sort of have a chat and a user profile and, and it will then either call you or WhatsApp you going forward. But it's pretty flexible. So you can, you can do several things to, to do that.

Demetrios Brinkmann [00:03:56]: Oh, this is super cool. And I'm thinking about all the times that I would want this or now I'm planning a trip with old friends, like a ski trip. And we all have families and it's all very difficult to organize every little piece of that. Right. And so thinking about an agent that can just help me know when the best time is for everyone, know where we are going to go and know like what our budgets are. Between all of us and all of that stuff, then that's already a feat in itself. But the flights feel like, how did you do that?

Marissa Liu [00:04:34]: We tested with a number of browser automation tools which. And they basically specify a workflow or through these tools you can specify a workflow that says navigate to this site, make these clicks. And we found that a lot of the tools out there in the market right now are quite unreliable. But we landed on skyvern, which kind of did a lot of the work out of the box. Everything from defining a workflow to navigate to booking.com or any travel agent site actually enter the search parameters that you require. So the flight dates, the destination and then Find the cheapest flight based on your prompt. And then finally the last mile. The most difficult part, which is what Charlie talked about, which is entering your payment details.

Marissa Liu [00:05:20]: It actually natively supports persisting your payment, your credit card details and safely kind of inputting them into the form for you. A lot of the hassle was actually covered by these great developers who thought the tool.

Demetrios Brinkmann [00:05:35]: How are you getting the credit card details from folks? Is that on that phone call?

Marissa Liu [00:05:39]: Either through the phone call or through WhatsApp. But for the purposes of this demo, we wanted to keep it secure and so we stuck with our own test credit card loaded with a pre defined, I guess like a fixed limit.

Charlie Cheesman [00:05:54]: So the ongoing solution to this, we were playing around with a few things of what could be done within the few days. Stripe is actually giving us access to. They set up a new tool, it's actually in private beta at the moment and we had a call with them at like 1am on the other day. They were super, super nice actually. So shout out to Stripe for that. But they're giving us access. They've just set up this thing called. It's a one time payments card.

Charlie Cheesman [00:06:20]: So you have your bank card that it links to, so the user would put that in and then they'll create a one time virtual card that will be used specifically for that transaction. It authorizes a certain payment amount and then after that the agent has access to that, uses it for the transaction, then afterwards that sort of disappears. So that would be like, oh, that's how we'll set it up.

Demetrios Brinkmann [00:06:40]: And are you double checking with the user before you're buying the stuff? Because I imagine it could get really dangerous when you're like, oh yeah, we bought you a first class ticket to Europe.

Charlie Cheesman [00:06:51]: Yeah, yeah, yeah.

Ana Shevchenko [00:06:52]: We are of course asking the user for, for confirmation. The more of a long term idea is to have a plan that will be displayed to the user in the app, but also like sent to the user as an email, as a WhatsApp message, whatever, for every single thing. So they can modify it or just confirm it and the agent will go off and execute all of these options. So we're planning to make it way more flexible and way more.

Demetrios Brinkmann [00:07:22]: I could see that I would want this just, you know, click Apple Pay, it pops up and it's working in the background and then you just click for Apple Pay and you go, all right, cool. I verify that I'm all right with spending 300 bucks for this flight or for putting the deposit on the hotel or whatever it may be.

Charlie Cheesman [00:07:42]: Yeah, yeah, yeah. Agent pops it in the group chat, like. Like this message. If it's all good, everybody likes the message. You've all paid. Like, you could imagine that that would be exactly how it goes. So it's. It's coming soon.

Demetrios Brinkmann [00:07:52]: So is this a company now or.

Charlie Cheesman [00:07:56]: Well, we have revenue, like, we have. We. We literally have people signed up. So, yes, I guess it is. So good. That's so good. Yeah.

Ana Shevchenko [00:08:05]: Done entirely within the hackathon, by the way.

Charlie Cheesman [00:08:07]: Yeah, yeah.

Demetrios Brinkmann [00:08:08]: Well, yeah. What was it, 24 hours, 48 hours.

Charlie Cheesman [00:08:10]: Yeah. So. So we had this thing internally. We were like, revenue by Tuesday, revenue by Tuesday. So we put up a post on. On. On the Sunday on Twitter. We put a bit of promotion behind it, went and talked to loads of people around here and people, like, are really happy to support that.

Charlie Cheesman [00:08:27]: They're like, okay, this is cool. Like, I can see the value of the use case. And then they signed up and you're like, oh, my God, somebody's actually signing up this thing. Oh, yeah. So the link is, yay, travel.

Demetrios Brinkmann [00:08:38]: Yay travel. All right, you got another user in me.

Charlie Cheesman [00:08:41]: Nice.

Demetrios Brinkmann [00:08:41]: All right, I like it. Is there anything else you want to hit on?

Charlie Cheesman [00:08:45]: I mean, shout out to the rest of the team? I think so. There were five of us, so Ferg and Alex as well. They're sort of around the conference somewhere. Yeah, the tech team was really awesome. There was another team, so we're all from a developer group called Unicorn Mafia. And, yeah, if you're a developer interested in hacking, then get in touch. We're always keen to hack.

Demetrios Brinkmann [00:09:03]: You're hiring? Yeah, yeah, yeah.

Ana Shevchenko [00:09:06]: You have to fill in the form.