BEAR: The Berkeley Event & Availability Recommender

A student-led capstone project at the University of California, Berkeley—conducted in partnership with 99P Labs at the Honda Research Institute USA—explored whether Large Language Models (LLMs) could meaningfully enhance trip and event planning. Focusing on the City of Berkeley, renowned for its vibrant cultural, academic, and community offerings, the team integrated LLM-driven recommendation engines with user calendars to build a system that dynamically filters, ranks, and recommends events based on individual preferences, schedules, and budgets.

‍

The work centered on ingesting diverse local event data, such as lectures, concerts, exhibits, and sporting activities, and then applying LLMs to match potential experiences with user-specific criteria. These criteria included location, scheduling, cost limits, and personal interests—like favorite music genres or culinary preferences. To streamline the process, the team integrated Python for the core logic, Pandas for data handling, and Gradio for an interactive user interface. LLMs processed prompts about available events and user constraints, returning curated lists that were further refined by calendar matching.

‍

Key findings showed that a more modular, “agentic” arrangement of smaller AI models performed comparably to prompting a single larger and more expensive LLM with the entire dataset. This suggests that flexible pipelines, which rely on multiple specialized models, can deliver similar quality without the higher operational overhead. User feedback confirmed that the prototype effectively simplified event discovery and helped them uncover locally relevant experiences they might otherwise have missed.

‍

Looking ahead, the team sees opportunities to expand beyond static databases, tapping into real-time APIs for a broader spectrum of live events. This next step would enable the system to adapt on the fly, adjusting schedules in response to last-minute changes or newly announced activities. Overall, the UC Berkeley team’s exploration illustrates how LLMs, when thoughtfully integrated into a pipeline of data collection, filtering, and calendar synchronization, can reshape local event planning to be more user-centric and efficient.