Applied AI System
AI Weather Predictor
End-to-end weather intelligence application combining data retrieval, feature engineering, sequence modeling, experiment tracking, Streamlit delivery, LLM summaries, and cloud-ready infrastructure.
Overview
AI Weather Predictor is a full technical product that forecasts the next 7 days of weather for a selected city. The system retrieves recent station data, runs a reproducible preprocessing pipeline, prepares a 30-day input sequence, generates multi-target forecasts with a trained Keras model, and explains the result in natural language.
The project goes beyond a notebook by connecting data collection, preprocessing, model inference, experiment tracking, cloud infrastructure, usage logging, and an interactive Streamlit application.
Problem
Weather forecasting is a practical technical problem because the data is noisy, time-based, location-dependent, and incomplete. A useful product needs data retrieval, feature engineering, model inference, reliability checks, and a user-facing explanation layer.
Solution
The application accepts a city, resolves nearby weather stations, fetches recent daily observations, and enriches the time series with calendar, cyclical, trend, weather-risk, wind-chill, and missing-value features. The final 30-day window is passed into a sequence model that predicts temperature, wind, precipitation, and snow values for the next 7 days.
The forecast table is then converted into a user-friendly weekly summary using an LLM interface.
Technical Highlights
- City-level data retrieval using geospatial lookup and weather-station data.
- Multi-target 7-day forecasting from a 30-day historical window.
- Feature engineering for time, trend, missing values, weather risk, and derived weather indicators.
- TensorFlow/Keras model artifact stored with supporting MLflow metadata.
- Streamlit product interface with forecast table, summary generation, and usage logging.
- Terraform resources for Google Cloud Storage and BigQuery foundations.
Architecture
The workflow separates the product into clear layers:
- User city input in Streamlit.
- Geocoding and nearby station lookup.
- Recent daily weather retrieval.
- Feature engineering and sequence creation.
- Keras model inference.
- Forecast table formatting.
- LLM-based explanation.
- Usage logging for later monitoring.
Future Improvements
- Add automated retraining and drift checks.
- Compare the Keras model with Transformer and Prophet baselines.
- Add CI checks for preprocessing and inference integrity.
- Containerize the app for managed cloud deployment.