Tripperoni

Chat Trip Recommendations
Sailing Trip Recommendation

Inspiration

Ourselves! One of our team members, Adam, always drives the discussions and organizes the trips - without him this trip and many more wouldn't have been possible! Since the group chats are at the heart of the idea exchange, containing the ideas, interests, opinions, and bouncing between websites and apps is just annoying, we wanted to make planning as seamless and convenient as possible.

What it does

Tripperoni is an AI travel agent completely residing inside your groupchat! You add him to your group chat, he analyzes your chat history, get every relevant information from you, and plans a trip. It consists of 3 main parts, with one planned, powered by the SkyScanner and the Gemini and OpenAI API: 1) Chatbot for group chats with an short- and long-term memory, that creates user profiles, can generate pictures and polls 2) Flight scanner via the SkyScanner API 3) Accommodation browsing agent, that goes through SkyScanner website and chooses the hotels that match the users preferences. Extra: Agent that books the accommodation, travel and event tickets.

How we built it

After a long ideation phase, we decided to split the work into four main parts: 1) Data base development to allow for persistent & continuous conversations and long + short memory influencing the context 2) Telegram Bot API integration + LLM Chatbot 3) SkyScanner Flight API integration 4) Agentic Web Scraping

Challenges we ran into

Getting the Long + Short term memory system to work was not trivial, but after a few iterations, we got it to produce meaningful results. Another important challenge we faced was to figure out when the Agent should enter the conversation or let the users talk to each other. Agentic Web Scraping remains an active research area - initially we couldn't get it to perform as well as we would've liked, facing long runtimes exceeding 30+ minutes for a single location, but managed to bring it down for our Proof-of-Concept that scrapes 20 locations in 3 minutes (9s per location).

Accomplishments that we're proud of

We're proud to see that our product manages to aggregate the needs and desires of all users within the chat by leveraging both state of the art LLM APIs and subtle tricks to improve chat quality and information retrieval. Each of us faced challenges - and in the end, we surpassed them all and were able to learn a lot.

What we learned

We learned to make use of the existing API of Telegram to bypass the need of developing our own messaging application. If this wouldn't have worked, we most likely wouldn't have had an MVP by the end of the hack. We have also learned to make LLM chats more cohesive, how to even use LLMs (some of never worked with LLM before) and how to integrate the Skyscanner API effectively.

What's next for Tripperoni

1) Make it completely End-to-End, by creating an agent that books the flights and hotels. There exist several challenges in regards to data security (sensitive credit card data to book the flight, two-factor authentication, etc.) 2) Integrations with users calendars (Google Calendar, Apple Calendar)