I started the food-delivery-time-prediction repository in 2026-09 and last updated it in 2026-09.
I call the application DeliveryETA. Its Streamlit frontend and FastAPI backend run as separate Docker-based services on Render. The live application is at https://food-delivery-time-prediction-2yntyjwau4y3ghbzhmy73r.streamlit.app/, the API is at https://food-delivery-time-prediction-0zo9.onrender.com, Swagger documentation is at https://food-delivery-time-prediction-0zo9.onrender.com/docs, and the health check is at https://food-delivery-time-prediction-0zo9.onrender.com/health. The deployed frontend connects to the backend and returns predictions from the trained pipeline.
My prediction flow is user → Streamlit frontend → POST /predict → FastAPI backend → preprocessing and feature pipeline → Random Forest Regressor → predicted ETA in minutes. The inputs cover rider age and rating; vehicle condition and type; restaurant-to-customer distance; traffic and weather; order type; multiple deliveries; festival conditions; city type; order timing; and preparation time. FastAPI uses Pydantic validation to reject invalid or unexpected fields.
I derive distance_km with Haversine distance between restaurant and delivery locations. My date and time features include order_day, order_month, order_weekday, is_weekend, order_hour, and order_time_bucket (Morning, Afternoon, Evening, or Night); prep_time_min represents order preparation or pickup time. Numerical and categorical features are processed with StandardScaler and OneHotEncoder.
My workflow starts with the raw Zomato dataset, then data cleaning, feature engineering, a processed dataset, a train/test split, preprocessing, model training, and evaluation. I compared Linear Regression, Random Forest, Gradient Boosting, and XGBoost; Random Forest had the lowest reported RMSE, so I selected it and saved the complete trained pipeline—not only the estimator—to models/best_model.joblib. The reported results were: Linear Regression, MAE 4.85, RMSE 6.08, R² 0.584; Random Forest, MAE 3.12, RMSE 3.92, R² 0.827; Gradient Boosting, MAE 3.61, RMSE 4.54, R² 0.768; and XGBoost, MAE 3.14, RMSE 3.93, R² 0.826. MAE is average absolute prediction error in minutes, RMSE penalizes larger errors more strongly, and R² is the proportion of target variance explained.
In the README example, my request returns {"predicted_minutes": 20.2}.
The example request uses age 28, rating 4.6, vehicle condition 2, 1 multiple delivery, Festival 0, distance 5.4 km, day 15, month 9, weekday 1, weekend 0, hour 19, and preparation time 12 minutes; its categorical values are Fog, High traffic, Snack, motorcycle, Metropolitian, and Evening. My EDA outputs cover the target delivery-time distribution, traffic and weather against delivery time, multiple deliveries against delivery time, and a correlation heatmap; the plots and model_metrics.json are under outputs/.
pip install -r requirements.txt
python src/preprocessing.py
python src/train.py
uvicorn src.api:app --reload --port 8000
streamlit run app/streamlit_app.py
pytest -v
docker compose up --build
I can run the frontend and API separately or start the complete application with Docker Compose. The local frontend is at http://localhost:8501, the API at http://localhost:8000, and Swagger at http://localhost:8000/docs. The API exposes GET /health, POST /predict, GET /docs, and GET /openapi.json. My Pytest suite covers health checks, valid predictions, validation errors, missing and extra fields, and response structure.
My documented limitations are that preparation-time information may not be available when an order is first placed, evaluation uses a single train/test split, model monitoring and data-drift detection are not implemented, and the free Render deployment may sleep after inactivity. Proposed improvements include cross-validation and hyperparameter tuning, model versioning, CI/CD, model and data monitoring, real-time traffic and weather features, and prediction-uncertainty estimates. The README does not provide the dataset source or size, split ratio, evaluation date, model hyperparameters, or live service performance.
I developed DeliveryETA as an end-to-end Machine Learning project to predict the estimated time required to deliver a food order, measured in minutes. I wanted to explore how Machine Learning could address a practical business problem: learning patterns from historical delivery data to predict delivery times for new orders instead of estimating them manually.
I framed this as a Supervised Machine Learning Regression problem because the target variable, Time_taken (min), is a continuous numerical value representing the actual delivery time in minutes. Food delivery platforms need reliable estimates, and a fixed-time estimate cannot properly account for the different conditions that affect delivery.
- I used a Zomato-style dataset containing historical order and delivery information. The listed features include delivery partner age and rating; restaurant and delivery location coordinates; weather and traffic conditions; vehicle type and condition; order type; festival information; and a final, incomplete entry beginning “City and”. Factors identified as affecting delivery time also include distance, restaurant preparation time, and other delivery conditions.
I trained models on Zomato_Dataset.csv, using order, rider, distance, traffic, weather, and preparation-time information to predict delivery time in minutes. My workflow covers data cleaning, missing-value handling, exploratory data analysis (EDA), feature engineering, model comparison, evaluation, API development, frontend integration, testing, and deployment. The project description does not state the dataset size or train/test split.
I used Zomato_Dataset.csv; the target is delivery time in minutes. The listed input features include rider age and rating, vehicle condition and type, delivery distance, traffic, weather, multiple deliveries, festival information, city, order date and time, and preparation time.
I used Haversine distance to calculate geographical distance in kilometres. From date and time information, I derived hour, weekday, month, a weekend indicator, and a time bucket. I processed numerical features with StandardScaler and categorical features with OneHotEncoder through Scikit-learn pipelines.
- I compared four regression algorithms:
Random Forest achieved my best recorded evaluation results: MAE of 3.12 minutes, RMSE of 3.92 minutes, and an R² score of 0.827.
My technology stack includes Python, Pandas, NumPy, Scikit-learn, XGBoost, FastAPI, Pydantic, Streamlit, Docker, Render, and Pytest. FastAPI provides a /predict endpoint, and Streamlit provides an interactive interface where users enter delivery details and receive predicted delivery times.
The project demonstrates an end-to-end machine-learning lifecycle, from data analysis and preprocessing through feature engineering, model comparison, evaluation, API development, frontend integration, testing, and deployment. The source ends with the incomplete phrase “Future imp”, so it does not specify future improvements.
- Linear Regression
- Random Forest Regressor
- Gradient Boosting Regressor
- XGBoost Regressor