Stardrift, an AI travel planning app, needed a system to transform freeform user queries like 'sfo-jfk' into compelling visuals. The challenge involved converting vague inputs into specific locations and then associating them with pleasing photos. The core problem was designing an AI system with a sense of aesthetic taste and practical functionality.

The solution involved a multi-stage process. First, LLMs were employed to parse user inputs and identify potential locations. Then, traditional software engineering facilitated the creation of a database linking places to images. Finally, human curation by the author ensured the selection of high-quality photographs, avoiding generic AI-generated results.

The project's success depended on defining a 'place' with a name and type, enabling the system to handle multiple destinations. A database was built by scraping Unsplash using an API, manually selecting the best images for each location. A fallback mechanism using geolocation provides photos for places not in the database, with manual additions to fill gaps.

This approach demonstrates how to integrate AI with traditional software engineering to create refined user experiences. The author highlights the importance of human oversight and curation in AI projects, specifically in the context of travel apps. The system's evolution continues with manual updates to address gaps and ensure image quality, suggesting an ongoing process of refining the app's visual appeal.