Beeline

System Design

Inspiration

Navigating with Google maps and existing navigation apps SUCKS especially for multi-modal uses where often subpar routes are recommend. Google Maps will suggest transit the whole way, or driving your personal vehicle, but wont suggest options like biking or ride-sharing to the nearest transit station or, if driving, wont suggest areas to park.

What it does

Beeline provides AI supercharged route recommendations that break your trip up into multi-modal legs where each leg is a different mode. This allows you to really personalize your navigation recommendations based on your personal preferences, especially for the first/last mile. Depending on the mode for each leg Beeline provides context aware suggestions: for example when driving Beeline will suggest convenient and safe places to park and when taking transit Beeline will take into account transfer wait times and suggest alternative modes as well; like bike-shares or ride-shares.

How we built it

We built a front-end where users can provide their route via natural human readable text. The front-end passes your route to the back-end which uses your LLM provider of choice and MCP integrated services to break your trip into multi-modal legs and plans each leg separately. The back-end returns the best three route options back to the front-end where they are rendered to the user. Eventually we will integrate more services and options so the user can export their trip legs back to the map provider of their choice and then use services like Google Maps on their phone to take each leg of the trip.

Challenges we ran into

We are not front end people so the entire front end. MCP integration as this was the first time we integrated an MCP server with an LLM. Time crunch Not having Anthropic credits so we couldn't use the existing Anthropic Google Maps MCP server So we had to pick a suboptimal one with Gemini and use Google Maps API

Accomplishments that we're proud of

Having a working demo by the deadline MCP server integration

What we learned

We learned about making a Frontend using Sveltekit as well as integrating MCP servers with Google Maps API.
This was also our first frontend and backend integrated project.
- We also realized that response time for these services are important and chose Gemini 2.5 Flash-Lite for the demo responses but we got better results with larger models like Gemini 2.5 Flash.

What's next for Beeline

We need to refine the routes perhaps based on segments since we are using Gemini 2.5 Flash-Lite
We need to add options for using different models
We also need to link the chatbot to the route planning and view different routes