This project was created at the Distributed AI Hack Berlin, held on November 14–15, 2025. For our analysis, we relied on the Pastry Prediction Data provided on Kaggle.
Bake Boost is the federated learning platform that lets bakeries collaborate on demand forecasting without sharing their sensitive data. Each bakery trains a model on their own sales history locally. These models share learnings with each other through Flower, a privacy-preserving federated framework, creating a collective intelligence.
Small bakeries face a daily dilemma: produce too much, waste food, money and labor at the end of the day. Or even produce too little and disappoint customers. Especially for small bakeries, this makes it extremely difficult to keep up with the large competition from chains. They lack the data science expertise and resources that large chains have. The result: massive food waste, higher costs, and inefficient staffing.
Comparative training on separate training data, used by the stores independently, show worse RMSE and MAE metrics. It is explained by insufficient amounts of data stored by some of the shops. Using federated learning allows the stores to accurately train the model while not directly sharing their data.
The data distribution of all datasets shows strong weekly seasonality and therefore shows similar correlation between features and target, which allowed easy integration.
- Python 3.9+ (Python 3.10+ recommended)
- Virtual environment activated
-
Clone and navigate to the project:
cd BakeBoost -
Activate virtual environment:
source venv/bin/activate -
Install dependencies:
pip install -r requirements.txt
# CPU
./submit-job.sh "flwr run . cluster-cpu"
# GPU
./submit-job.sh "flwr run . cluster-gpu" --gpu
# View running jobs
squeue -u $USER
# View job output
cat ~/logs/job_*.outTest with 5 simulated bakeries on your local machine:
flwr run . local-simulationBuilt with Flower - The Friendly Federated Learning Framework