AI suggested orders

Currently in the food to go industry, wastage of short shelf-life food, like sandwiches, wraps, salads, etc., can be in excess of 5% of its revenue. This is costly to both retailers and manufacturers who fail to accurately predict how much they will sell each day. The FtG industry in the UK alone is worth ~£22B and therefore the wastage problem is approximately £1.1B per year.

Through our company Mezze Software Ltd, over the past 3 years we have processed over £120 million of orders of Food to Go products through our existing B2B eCommerce platform. Consistently our clients (sandwich manufacturers like Samworth Brothers, Simply Lunch, Real Wrap Co. Tiffin Sandwiches) have voiced to us the challenge of predicting orders and the cost of this waste to them and their customers as well as concern over the environmental impact of this waste. Some of our customers have already looked at non-industry specific software to solve this issue with very little success.

Our project aims to prove that a machine learning model, when provided the right data, could predict orders of sandwiches more accurately than any human could, and at a cost that is affordable for both large and small manufacturers. Our model would aim to be at least 96% accurate at correctly predicting orders of sandwiches by the end of this project, improving on the current industry waste levels. A 1% increase in accuracy would save over 300,000 sandwiches being wasted for just one of our clients, saving over £630,000\.

To achieve this accuracy we aim to test various machine learning algorithms for a base level of accuracy. After this we would iterate over the base machine learning model experimenting with variables such as historic waste, weather, seasonality, geographical location, local events and ingredients to improve accuracy. We will also build an interface for the retailers ordering the sandwiches to interact with, to provide feedback of accuracy of the orders so that we can further fine tune the model.

We will collaborate closely with our clients on this project. With their input we can progressively trial the new product in the real world with existing customers and monitor accuracy but also keeping in mind the cost of the end product. The trial can then transition to replacing human ordering for those real customers.