AirGuide Agent

Inspiration

The idea for AirGuide Agent was inspired by a real-world need to help travelers with specific accessibility requirements find suitable Airbnb listings more easily. While platforms like Airbnb offer filters, they often fall short of capturing accessibility feedback found in reviews. We wanted to create an intelligent agent that could surface relevant and inclusive options by analyzing real guest experiences.

What it does

AirGuide Agent helps users:

• Find Airbnb properties in a specific city that have high reviews.
• Identify listings that may offer accessibility features based on review content.
• Summarize key property attributes and guest impressions using natural language understanding.
• Act as a helpful travel companion that recommends inclusive stays.

How we built it

We used:

• Databricks and SQL UDFs to filter Airbnb properties with strong reviews.
• A custom Python function tool to scan reviews for accessibility-related keywords.
• Delta Sharing to connect external datasets securely.
• Databricks Agents to orchestrate interaction between SQL, Python, and the user via a natural language interface.

Challenges we ran into

• Delta Sharing catalogs are read-only, so we couldn’t create functions directly within shared schemas.
• Passing arrays of strings to Python functions required careful handling of serialization.
• Mapping free-form text (reviews) to structured insights was challenging and required thoughtful keyword matching logic.
• Designing intuitive interactions between SQL and Python tools within the agent's framework took iteration.

Accomplishments that we're proud of

• Created a working agent that connects SQL and Python tools seamlessly.
• Enabled accessibility analysis directly from guest reviews — something not available on most platforms.

What we learned

• How to integrate Delta Sharing data into Databricks workflows.
• The strengths and limits of function tools in SQL vs. Python within Databricks Agents.
• How to manage data types and outputs to avoid common errors (e.g., JSON serialization issues).
• Importance of agent system prompts and function design for quality interactions.

What's next for AirGuide Agent

• Add support for multi-language review analysis.
• Integrate LLM-based summarization to offer pros/cons of each listing.
• Expand keyword sets using embeddings or fine-tuned models for accessibility classification.
• Enable users to filter listings based on custom preferences (e.g., pet-friendly, elevator access, etc.).
• Deploy the agent as part of a larger travel assistant platform or integrate it with messaging apps.

Built With

• langgraph
• mlflow
• mosaicagentframework