WeatherWear

AI-powered outfit recommendations based on real-time weather data.

Built as the final project for CSE 150A (Introduction to Artificial Intelligence: Probabilistic Reasoning and Decision-Making) at UC San Diego.

What It Does

WeatherWear is a utility-based AI agent that recommends what to wear based on the weather. Select any city in the world and the agent will:

• Fetch real-time and forecasted weather data client-side (via the free Open-Meteo API)
• Classify weather conditions using a Random Forest model trained on historical Seattle weather (2012-2015)
• Recommend a full outfit (top, bottom, outerwear, footwear, accessories) with a conversational, emoji-rich explanation
• Show a weekly view (Sunday-Saturday) with outfit summaries for each day

The entire system runs without any paid APIs or LLMs.

Live Demo

weatherwear-kpgz.onrender.com (free tier — first load may take ~30s to wake up)

Features

• City search with smart country detection (type "Japan" and get Tokyo, Osaka, Kyoto, etc.)
• Geolocation support ("Use my location")
• Native language names for non-Latin cities (Tokyo → 東京, Cairo → القاهرة, Munich → München)
• Clickable weekly forecast — click any day to see the full breakdown
• Past vs. future tense — past days use retrospective language, future days use forecast language
• Dynamic weather themes — background changes based on predicted conditions (sun, rain, snow, etc.)
• Celsius / Fahrenheit toggle
• Methodology section explaining the full decision pipeline on the page itself

How It Works

 Browser                                          Server
 ──────────────────────────                        ──────────────────────────

 Open-Meteo API (free)                                    
   weather forecast data                                  
         │                                                
         ▼                                                
 4 features per day             POST /api/recommend       
 (precip, temp, wind)  ──────────────────────────▶  Random Forest Classifier
                                                    (400 trees, scikit-learn)
                                                             │
                                                             ▼
                                                    Rule-Based Outfit Policy
                                                             │
                               ◀─────────────────────────────┘
                               JSON: label, outfit, explanation

Model Performance

Weather	Precision	Recall	F1
Rain	0.97	0.91	0.94
Fog	0.87	0.96	0.91
Snow	0.89	0.89	0.89
Drizzle	0.82	0.91	0.86
Sun	0.76	0.92	0.83

PEAS Analysis

Component	Description
Performance	Accuracy of weather classification and appropriateness of outfit recommendations
Environment	Real-world weather for any city (Open-Meteo API) + historical Seattle data for training
Actuators	Outfit recommendations (top, bottom, outerwear, footwear, accessories) with natural-language explanations
Sensors	Precipitation (mm), temp max/min (°C), wind speed (m/s) from Open-Meteo forecasts (fetched client-side) or manual CLI input

Tech Stack

• Backend: Python, FastAPI, scikit-learn, Jinja2
• Frontend: HTML, CSS, vanilla JavaScript (calls Open-Meteo directly from the browser)
• ML Model: Random Forest (scikit-learn) trained on Seattle Weather dataset
• APIs: Open-Meteo (weather + geocoding), Nominatim (reverse geocoding)
• Hosting: Render (free tier)

Run Locally

# Clone and set up
git clone https://github.com/joshleh/WeatherWear.git
cd WeatherWear
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

# Start the web app
uvicorn weatherwear.webapp:app --reload --port 8010

Then open http://localhost:8010.

CLI (offline, no network needed)

# Train the model
weatherwear train --data seattle-weather.csv

# Demo from a date in the dataset
weatherwear demo date --date 2012-01-01

# Demo with manual weather inputs
weatherwear demo manual --precipitation 3.2 --temp-max 11 --temp-min 5 --wind 6.5

Project Structure

src/weatherwear/
    agent.py             Sense → Think → Act orchestration
    model.py             Random Forest training, persistence, inference
    outfit.py            Rule-based outfit policy
    explain.py           Template-based explanation engine (no LLM)
    weather.py           City search + reverse geocoding (Open-Meteo / Nominatim)
    webapp.py            FastAPI web application (serves UI + /api/recommend)
    cli.py               Command-line interface (offline demos)
    types.py             Shared data types
    country_cities.json  Pre-computed city data for 50+ countries

web/
    templates/           HTML + client-side JS (fetches weather directly from Open-Meteo)
    static/              CSS styles

seattle-weather.csv      Training dataset (Seattle 2012-2015)
render.yaml              Render deployment blueprint

Limitations

• The model is trained on Seattle weather (2012-2015). Cities with extreme climates may get unusual classifications, though the outfit policy still uses raw feature values as a fallback.
• The explanation engine is template-based and deterministic — not LLM-generated. This keeps the project 100% free with no API keys.
• Open-Meteo provides up to 7 days of forecast data; accuracy naturally decreases further out.

License

This project was created for educational purposes as part of CSE 150A at UC San Diego.

Acknowledgments

• Dataset: Seattle Weather (Kaggle)
• Weather API: Open-Meteo — free, no API key required
• Geocoding: Nominatim / OpenStreetMap — free, open-source