AI TravelMate

Inspiration

Our team is pretty diverse as we came together representing 4 different companies and corresponding domains at the start of the hackathon. We were inspired by the bright initiative data sets and our own experiences with our unique wants and needs when planning a stay away from home to create a friendly travel agent AI TravelMate who can minimize the angst associated with planning a trip and empower us to consider choices beyond those enabled by a usual search filter..

What it does

We’re building an AI-native, “Kayak-like” travel assistant platform that consolidates data from multiple sources—including Airbnb, Google, and Booking.com—into a single interface.

The system helps users find the best stay options by interacting with an AI trip consultant agent, which guides them through a structured conversation:

Greet the user Ask for destination Ask how many people are traveling Ask for budget, special requirement.

From this chat, the agent extracts structured constraints (e.g., location, price, number of guests) to drive intelligent search and recommendation.

How we built it

Under the hood, this is a multi-agent system with the following architecture:

Conversational interface: A front-facing agent collects user input in natural language and translates it into structured constraints.

Schema discovery: Simultaneously, the system retrieves metadata (e.g., table names, schemas) from Genie conversation API

Genie conversational API: We use Genie to interpret user constraints in the context of available schema and generate appropriate SQL queries.

Front end integration: Genie’s output is passed sql query back to a query execution function (UDF), which generates and executes the final SQL query to retrieve the best-matching stays.

Core Components:*

LLM Foundation model (Meta Llama 3.3 70B Instruct):

We use built in foundation model for our front desk agent with comprehensive prompt instruction in travel-related queries, emphasizing natural conversation, itinerary planning, and contextual memory to guide users through complex booking journeys. Data Connectors UDF:

We built connectors to execute query to ingest and aggregate structured datasets from Airbnb and hotels, enabling rich metadata querying (e.g., amenities, reviews, price, location, availability).

Prompt Engineering for Safety & Role Integrity:

We developed a hardened system prompt that maintains the assistant’s role, resists injection attempts, and ensures secure, user-friendly interactions without exposing internal rules.

Genie conversation API: We use this to get data schema and metadata from unity catalog, sharing with our front-desk agent. Also based on constraints summarized from front-desk agent, constructing the sql query.

ChatBot:

AI TravelMate is deployed as a responsive, mobile-friendly chat designed for intuitive input Privacy & Security:

AI TravelMate is built with a strong privacy-first approach. No personal user data is stored or used for training, and the agent is designed to reject any attempts to extract system instructions or manipulate its behavior.

Challenges we ran into

Creating session in UDF we ran into problem to create spark session in UDF

**We got difficulty on setting up connection between genie and chatbot.

Data Normalization Across Platforms:

Problem: Airbnb and hotel datasets follow different schema structures, terminologies, and metadata formats

Handling Ambiguous or Vague Queries:

Problem: Users may submit open-ended questions like “Where should I stay in Europe in July?” which lack specific filters.

** Balancing LLM Creativity with Factual Accuracy**

Problem: The language model may generate plausible but incorrect or outdated listings if not grounded in real data.

** Use Vector DB and RAG for post-processing **

** Prompt Injection and Role Exploitation Risks**

Problem: Users may try to bypass the assistant’s guardrails by injecting meta-instructions

Latency in Data Retrieval

Problem: Real-time querying of large accommodation datasets may introduce noticeable delays

** Geolocation and Local Context Awareness**

Problem: Users may request “close to the beach” or “near downtown” without specifying exact locations.

User Interface Constraints

Problem: Users may expect listings with rich metadata (images, amenities, reviews) that are challenging to present in a compact chat interface

Cost Management

Problem: Need a way to track usage at scale

Hackathon limits

Problem: we will ultimately need a way to scale outside current limits

Accomplishments that we're proud of

**- register serveral functions in unity catatlog

- Having a working product! - Live Chat Interface with AI assistant

- Integrating AIrBNB and Hotel Data

- Promp-driven natural language filtering

- Conversational flow for booking recommendations

- System prompt security layer

What we learned

- We were able to achieve a lot even as a group of random strangers who came together for the hackathon and could have done even more if we had been able to search and connect before the show.

Also

- LLMs need guardrails

- LLMs need a little effort to create magic with structured data

- Designing Conversational UX is more complex than we had thought

- Simple queries like asking for recommendations near a location may require non-trivial computations if we cannot rely on the LLM's expertise

- Limited time forced us to focus on the essentials

- AI projects need to design safety to prevent being tricked

- Team coordination matters - we could accomplish what we set out to do by dividing responsibilities

What's next for AI TravelMate –

Smart Recommendations

The AI can suggest accommodations based on your travel style, previous interactions, and preferences—whether you're a solo traveler, a remote worker, or a family on vacation.

Interactive Itinerary Building

Combine accommodation search with nearby attractions, restaurants, and transport options to build a seamless travel plan, all through a single interface.

** Use Vector DB and RAG for post-processing **

Built With

• brightinitiativedatasets
• databricks
• genie
• llama
• playground
• python