We augment the purchasing and procurement process to improve accuracy and reduce food waste through the use of Artificial Intelligence. With a strong faith in data-driven decisions, we eliminated the subjectivity and potential human error in the traditional purchasing and procurement process.
What it does
From each store/client’s past history, a forecast is made to give advice on what is the optimal amount of food to order. As compared to the traditional method of consulting or decision-making by an experienced chief, ambiguity is removed in the procurement of raw materials. In exchange, an accurate purchasing can be done to minimize waste and cost while keeping inventory fresh.
How we built it
We tried several time-series forecasting architecture such as Long Short Term Memory (LSTM), Convolutional Neural Network 1D (Conv1D), CNN-LSTM, Gated Recurrent Unit (GRU) and Recurrent Neural Network (RNN). Based on the performance, we decided that RNN is the most suitable in terms of the loss obtained. We then tuned the hyperparameters to further reduce the loss and optimize the model.
Challenges we ran into
This is our first time working with a time-series forecast and there were some challenges learning about the architecture used for it. Preparing the data was time-consuming as we try to prepare data which is representative of what is being sold.
Accomplishments that we're proud of
The first time we obtained a prediction was a memorable one and we were able to test across the different architecture to find the best architecture for the task. Being able to reduce the loss by tuning the parameters was another success for us.
What we learned
We learned a great deal about the different architecture for forecasting and the hyperparameters involved in it. We were also able to gain some valuable experience preparing the data for model fitting.
What's next for BuyFresh
Further work can be done by including more features or implement some feature engineering to be able to account for freshness of the purchases better.