CitiBike Agent

• 

  Flow diagram

🚲 Inspiration Downtown Manhattan’s Citi Bike stations are constantly imbalanced—some completely empty, others overflowing. This frustrates users and forces operators to reactively rebalance bikes. We were inspired to build an AI-powered assistant that forecasts these imbalances and suggests proactive actions before problems arise.

⚙️ What it does CitiBike Agent is a smart assistant that:

Monitors live bike station data in real time

Pulls historical trip, weather, and time-of-day data using BigQuery

Forecasts upcoming shortages or surpluses using AI

Suggests optimal bike truck roll movements to rebalance the system

Communicates with users via a Firebase-hosted chatbot and map UI

🛠️ How we built it Backend: Google Cloud Vertex AI Agent Builder + BigQuery

Live Data: Real-time station status ingested from Citi Bike’s feed into SQL

Forecasting: Gemini-powered sub-agents use weather, station, and historical trip data

Frontend: Firebase web app with embedded map and chat agent

Deployment: Agents deployed via GKE/Cloud Run, served securely with service account tokens

🧱 Challenges we ran into Making Google AI Agents access and process BigQuery data in a structured way

Combining weather, geospatial, and trip data into useful forecasting prompts

Syncing Firebase frontend with live chat and real-time map updates

Deploying reliably with proper credential management across cloud services

🏆 Accomplishments that we're proud of Built a multi-agent AI system that interprets live and historical data to give meaningful operational suggestions

Developed an intuitive, real-time dashboard and chatbot to interact with the agent

Deployed a functioning end-to-end system that showcases the power of AI in solving real-world transit problems

What we learned How to orchestrate Gemini-based agents for forecasting and monitoring tasks

Best practices for integrating Vertex AI, BigQuery, and Firebase into one seamless experience

Importance of data storytelling: clean visualization and summaries helped surface useful insights from messy data

What's next for CitiBike Agent Add a mobile-friendly UI for riders and city planners

Train the forecast model on more granular event data (e.g., traffic or subway outages)

Integrate optimization algorithms to automate truck roll planning

Scale to other bike-sharing systems and micromobility networks globally

Built With

• adk
• bigquery
• citinyc
• cloudrun
• cron
• firebase
• linux
• python
• sql
• statmodels
• weatherapi