BIXI-MLOps

UrbanBikeML: A Reproducible MLOps Pipeline for Smart Mobility Analytics

---

1. Overview

This project demonstrates an end-to-end MLOps pipeline for predicting trip durations in Montreal's BIXI bike-sharing system. Leveraging real-world open datasets from BIXI Montréal, we apply machine learning to forecast how long a bike trip will take based on contextual and spatial features.

The pipeline includes model training, evaluation, deployment via FastAPI, and orchestration using Prefect — all designed to be modular, reproducible, and deployable.

---

2. Why this project?

Urban mobility services like BIXI are essential for sustainable and efficient cities. Predicting bike trip durations helps:

• Enhance user experience via better time estimation
• Improve bike redistribution operations
• Support incentive structures and dynamic pricing
• Integrate better into multimodal smart city systems

This project demonstrates how MLOps can be used in real-time transport solutions with public datasets.

---

3. Dataset

The data used in this project comes from BIXI Montréal’s official Open Data Portal. Specifically, the dataset for the 2023 season was used:

• Time range: April–November 2023
• Trip-level data: start time, end time, duration, stations, distances
• Station metadata: coordinates, arrondissement
• Format: CSV

We engineered features such as:

• trip_distance_km, hour, day_of_week, month, is_weekend
• start_arrondissement, end_arrondissement
• station_popularity, and more

---

4. Problem Statement

The goal is to build a machine learning model that predicts:

How long a given BIXI bike trip will take (in minutes), based on trip context

Using features such as:

• Distance between stations
• Temporal info (hour, day of week, month)
• Popularity and location of start/end stations
• Spatial coordinates and administrative zones

---

5. Project Architecture

BIXI-MLops/
├── app/
│   ├── main.py                   # FastAPI app exposing the prediction API
│   ├── model_loader.py          # Helper function to load the trained model
│
├── imgs/                        # Visual outputs and data visualizations
│   ├── ChatGPT Image Jul 28, 2025, 01_32_35 PM.png
│   ├── Top Origin-Destination Pairs.html
│   ├── start_station_heatmap.html
│
├── model/                       # Trained model artifacts (e.g., .pkl files)
│   └── best_model.pkl           # Output of Prefect pipeline model training
│
├── notebooks/                   # Jupyter notebooks for step-by-step development
│   ├── 01_data_cleaning.ipynb
│   ├── 02_EDA.ipynb
│   ├── 03_feature_engineering.ipynb
│   ├── 04_model_training.ipynb
│   ├── 05_orchestration.ipynb
│   └── mlruns/                  # MLflow experiment tracking output
│
├── pipeline/                    # Prefect flow and task definitions
│   ├── flow.py                  # Prefect flow logic (ETL + ML pipeline)
│   ├── tasks.py                 # Individual Prefect tasks
│   ├── utils.py                 # Helper functions used in pipeline
│   └── __pycache__/             # Python cache (auto-generated)
│
├── .gitignore                   # Git ignored files
├── Dockerfile                   # Docker container specification
├── requirements.txt             # Python dependencies
└── README.md                    # Project documentation (you’re writing this now)

---

6. Tech Stack

• Python 3.12
• JupyterLab
• Scikit-learn, Pandas, Joblib
• FastAPI (for API)
• Prefect 3.0 (for orchestration)
• Docker (for deployment)

---

7. Model Evaluation

Final trained model (Random Forest):

• RMSE: 8.69 minutes
• R² Score: 0.46

This means the model captures nearly half of the variation in trip duration — with potential for improvement using weather/user-type data.

---

8. Run the Project Locally

8.1. Clone the repository

git clone https://github.com/your-username/bixi-mlops-project.git
cd bixi-mlops-project

8.2. Install dependencies

python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

8.3. Run the pipeline (via Jupyter or Prefect)

Using Jupyter:

# Open notebooks/train_model.ipynb
# Run all cells to train and save model

Or via Prefect:

prefect deployment build pipeline/training_flow.py:BIXI-ML-Training-Flow -n dev
prefect deploy
prefect agent start

8.4. Launch the API

uvicorn app.main:app --reload

Then visit: http://127.0.0.1:8000/docs

9. Docker (Optional)

Build the image

docker build -t bixi-mlops-app .

Run the container

docker run -p 8000:8000 bixi-mlops-app

10. API Reference

POST /predict Input example:

{
  "trip_distance_km": 2.5,
  "hour": 17,
  "day_of_week": 5,
  "is_weekend": 0,
  "start_lat": 45.508,
  "start_lon": -73.561,
  "end_lat": 45.511,
  "end_lon": -73.554,
  "start_station_popularity": 10,
  "avg_trip_duration_from_station": 12.3,
  "start_arrondissement": "Ville-Marie",
  "end_arrondissement": "Le Plateau-Mont-Royal",
  "month": 7
}

Response:

{
  "predicted_trip_duration_min": 18.03
}

11. Limitations & Future Work

• Weather and user-type not yet integrated
• No dynamic model re-training pipeline
• Does not include station-level rebalancing strategies
• Can be extended with a dashboard or React frontend

12. Acknowledgements

• BIXI Montréal for releasing such a great open dataset.
• Open-source contributors to FastAPI, Prefect, Scikit-learn, and the Python community.