https://github.com/cnoret/getaround-pricing
Predict optimal rental prices and analyze delay patterns on Getaround. FastAPI prediction API + Streamlit dashboard, fully Dockerized with MLflow tracking.
https://github.com/cnoret/getaround-pricing
data-science docker fastapi getaround machinelearning mlflow mlops price-prediction python scikit-learn streamlit
Last synced: about 16 hours ago
JSON representation
Predict optimal rental prices and analyze delay patterns on Getaround. FastAPI prediction API + Streamlit dashboard, fully Dockerized with MLflow tracking.
- Host: GitHub
- URL: https://github.com/cnoret/getaround-pricing
- Owner: cnoret
- License: mit
- Created: 2025-07-14T17:34:48.000Z (12 months ago)
- Default Branch: main
- Last Pushed: 2026-06-16T15:18:38.000Z (13 days ago)
- Last Synced: 2026-06-16T16:05:07.864Z (13 days ago)
- Topics: data-science, docker, fastapi, getaround, machinelearning, mlflow, mlops, price-prediction, python, scikit-learn, streamlit
- Language: Python
- Homepage: https://cnoret-getaround-dashboard.hf.space/
- Size: 6.21 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Getaround — Price Prediction Dashboard & API
[](https://cnoret-getaround-dashboard.hf.space/)
[](https://cnoret-getaround-API.hf.space/)
[](https://github.com/cnoret/getaround-pricing/actions/workflows/ci.yml)
End-to-end ML project on Getaround car rental data - a price prediction API, an interactive analytics dashboard, and a full training pipeline with MLflow experiment tracking.
**Datasets:** pricing (~4,800 car listings · 13 features) · delay analysis (rental records with checkout times and consecutive booking gaps)
---
## Features
### Delay Analysis
- Interactive histogram of checkout delays with sentinel-value filtering (126 min encoding for >2h delays)
- Adjustable buffer threshold to simulate minimum time between consecutive rentals
- Business impact metrics: revenue at risk, blocked rentals, conflict resolution rate
- Delay breakdown by check-in type (Connect vs. mobile) with box plots
### Pricing Analysis
- Dataset KPIs: car count, average price, average mileage, average engine power
- Distribution charts for mileage and daily rental price
- Scatter plot: rental price vs. mileage colored by fuel type
- Average price by car type and by fuel type
### Price Prediction
- Form wired live to the FastAPI `/predict` endpoint
- Instant price estimate with delta vs. platform average
- Model metrics (MAE, RMSE, R²) in an expandable section
### Prediction API
- `POST /predict`: accepts one or multiple cars, returns predicted daily prices in euros
- `GET /health`: health check endpoint
- Pydantic v2 validation: field constraints + `Literal` types for all categoricals
- Interactive Swagger UI at `/docs`
---
## Tech Stack
| Layer | Libraries |
|--------------------|---------------------------------------------------------------------------|
| Data | Pandas |
| ML | Scikit-Learn (RandomForest, Pipeline, ColumnTransformer) · Joblib |
| Experiment tracking| MLflow |
| API | FastAPI · Pydantic v2 · Uvicorn |
| Dashboard | Streamlit · Plotly |
| Infra | Docker · Docker Compose |
| CI | GitHub Actions |
| Deployment | Hugging Face Spaces |
---
## Quick Start
```bash
docker compose up --build
```
| Service | URL |
|------------------|---------------------------------------------------|
| Dashboard | [localhost:8501](http://localhost:8501) |
| API + Swagger UI | [localhost:8001/docs](http://localhost:8001/docs) |
| MLflow UI | [localhost:5001](http://localhost:5001) |
Docker Compose trains the model first (`condition: service_completed_successfully`), then starts the API and dashboard. Inter-service communication uses the `API_URL` environment variable.
### Run locally without Docker
```bash
git clone https://github.com/cnoret/getaround-pricing.git
cd getaround-pricing
pip install -r api/requirements.txt -r dashboard/requirements.txt
# Terminal 1 — API
uvicorn api.main:app --reload --port 8001
# Terminal 2 — Dashboard
streamlit run dashboard/app.py --server.port=8501
```
---
## Project Structure
```text
.
├── api/ # FastAPI prediction API
│ ├── main.py
│ ├── requirements.txt
│ └── tests/
│ └── test_api.py
├── dashboard/ # Streamlit dashboard
│ ├── app.py
│ └── requirements.txt
├── data/ # Raw datasets
│ ├── get_around_delay_analysis.csv
│ └── get_around_pricing_project.csv
├── ml/ # Model training
│ ├── model/
│ │ └── model.joblib
│ ├── model_training.py
│ └── requirements.txt
├── notebooks/ # Exploratory Data Analysis
│ ├── eda_delay.ipynb
│ └── eda_pricing.ipynb
├── .github/workflows/ # CI
│ └── ci.yml
├── Dockerfile.fastapi
├── Dockerfile.dashboard
├── Dockerfile.training
└── docker-compose.yml
```
---
## Notebooks
| Notebook | Description |
|--------------------------------------------------|------------------------------------------------------------------|
| [eda_delay.ipynb](notebooks/eda_delay.ipynb) | Delay distribution, sentinel value analysis, conflict simulation |
| [eda_pricing.ipynb](notebooks/eda_pricing.ipynb) | Pricing EDA, feature correlations, model selection rationale |
---
## Results
| Metric | Value |
|--------|---------|
| MAE | 10.68 € |
| RMSE | 16.73 € |
| R² | 0.734 |
Random Forest with StandardScaler on numeric features (mileage, engine power) and OneHotEncoder on categoricals (brand, fuel, color, car type), evaluated on a 20% holdout set (random_state=42). Key predictors: engine power, car type, and model brand.
---
## Tests
```bash
pip install -r api/requirements.txt pytest httpx
pytest api/tests/ -v
```
11 tests covering `/health`, `/`, and `/predict`, including Pydantic validation errors (negative mileage, invalid fuel, missing fields). Model is mocked, no `.joblib` required. Also runs on every push via GitHub Actions.
---
## Retrain
```bash
pip install -r ml/requirements.txt
python ml/model_training.py
```
Metrics and artifacts are tracked in MLflow (`./mlruns`).
---
## License
MIT — [LICENSE](LICENSE)