OyaGo: Asking for Bus Directions in Lagos, Where No Map Knows the Routes

Lagos is a sprawling city in Africa with over 15 million inhabitants. Most residents navigate the city using an unstructured transport system composed of Danfo minibuses, BRT buses, keke motorized tricycles, and okada motorcycle taxis. Navigation relies on word-of-mouth directions which can be difficult for newcomers to comprehend.

Google Maps provides drive directions but does not offer guidance on which Danfo to board, conductor shouts, fare costs, or landmarks to look for. This lack of information often discourages people from venturing out, causing them to remain at home, take expensive ride-hailing services, or depend on friends for assistance.

OyaGo is a web application that addresses this problem by providing bus-by-bus directions from A to B in plain English or Nigerian Pidgin. Users can input their starting and ending points and receive a list of numbered steps detailing where to board, what to expect from the conductor, where to alight, and landmarks to guide them.

Users can also save their frequently visited locations such as home, work, and church, then simply request directions to these places. Additionally, users have the option to contribute their own routes, which enhances the app for the entire community.

OyaGo's architecture consists of a React Progressive Web App (PWA) for the phone interface, a FastAPI backend, Backboard for memory storage and retrieval, and an open-weight model developed by Google, specifically the Gemma-3-27b-it model. The model retrieves relevant route notes using RAG (retrieval-augmented generation) and generates directions based on those notes.

The mobile app receives these directions in a clear, conversational manner, either in Pidgin or plain English, matching the user's input style. All interactions are designed to be concise, allowing users to quickly receive navigation instructions at bus stops and continue their journey without relying on a phone screen. The system is built to be cost-effective, allowing it to remain free for users.

While it currently relies on a hosted open-weight model, the architecture is designed to be adaptable, enabling the model to run on various platforms, including offline, thereby ensuring accessibility regardless of internet connectivity.