Algeria Weather Forecaster

For the Advanced Machine Learning course of my master's, our team of four built a climate forecasting pipeline for Algeria. Rather than commit to a single model, the system lets a Q-learning agent choose, for each calendar month, whether to trust LSTM, SARIMA, Ridge or Prophet. We sketched the idea in December 2025 and built it in January 2026.

What I built

A teammate wrote the first Q-learning agent; I rewrote it into the router the system uses. I pretrained it on historical backcasts, so it learns from months where the true values are known, and used it to produce forecasts up to 2040. I also built the conversational side: a RAG chat in Streamlit, with ChromaDB, e5-small-v2 embeddings and Groq, that answers questions from the forecasts and climate statistics, and the generator for the PDF climate reports. My teammates built the preprocessing, the SPI and SPEI drought indices, the linear model, and the first SARIMA and LSTM models.

What I took from it

The model router was the idea I found most interesting: an agent deciding, month by month, which model to trust. It was also my first real RAG system, built for a practical reason: instead of reading through forecasts and reports by hand, you can ask an assistant about Algeria's weather and get the answer.