Food Banks in Latin America are investing 80% of their effort in procuring industrialized food that only fulfills less than 20% of their beneficiaries’ nutritional needs, mainly due to cost-efficiency reasons. And yet, in Latin America 80% of food waste happens in farms. And while the fruits and vegetables that are wasted could easily fulfil 80% of our beneficiaries’ nutritional needs, today they represent less than 20% of the Food Banks’ stock and rescue efforts.
Our goal with this project is to apply artificial intelligence to flip this balance, by applying machine learning and deep learning to develop a form of predictive logistics that can lower the costs of rescuing food from farms, and at the same time maximize the nutritional value of every order delivered by Food Banks to their beneficiaries. Using images from farmers when they begin their harvest, combined with historical data on average waste volumes for that particular product, we believe we can anticipate the volume of waste they will have at the end of the harvest. If we combine them with images from empty spaces in food transportation vehicles (trucks and utility vehicles) and powerful geolocation technology, we believe we can develop a system where we utilize the empty space in the vehicles and, based on their planned route and/or proximity to beneficiaries, assign them trips to pick up and deliver the food automatically. In addition, if we request from soup kitchens to share pictures of their last three supermarket purchases in a given period, we can cross the data with the rest of the soup kitchens in the vicinity and create nutrition profiles of each soup kitchen (i.e. the combination of vitamins, mineral, carbs and protein that they normally consume). And based on these profiles, we should be able to assign to them food donations that contain the nutrients that they need most based on their consumption patterns, as opposed to those that have maximum nutritional value on an absolute scale.
Our objective is to maximize the volume of fruits and vegetables rescued from farms and delivered to our beneficiaries. In order to do so, we would apply artificial intelligence in three stages: In farms (1): using image recognition and deep learning, we would ask farmers to send us pictures of their harvest, where both products harvested and wasted are visible (e.g. when fruits are left unharvested on the ground or on the plants). Our system would learn the types of fruits and vegetables contained in the picture, the average volume they occupy and the percentage of the harvest that is likely to be left to waste in the whole farm. Using satellite images and surveys to farmers, we would detect crops and their volume to estimate waste quantity. Using image recognition and deep learning, we would ask farmers to scan fruits and vegetables left unharvested on the ground in order to detect their quality and forecast the final available quantity. Once we know how many products are there to be rescued in a given farm, we will understand how many cubic meters of transportation space we will need to transport the food. That's when our registered vehicles (2) come into play. Using a system where drivers can indicate the idle capacity in their trucks (either planned or actual) and combined with image recognition and deep learning, we would ask drivers who have idle space in their vehicles to send us pictures of their cargo space. The system would then learn to calculate the amount of space available in each vehicle, and whether it has refrigeration capacity. We would then obtain from drivers their destination information, in connection with the trip they are taking or planning to take that day or week. If the trip is scheduled for another time or day, we would ask drivers to simply list the cargo (product and units) along with a picture of their empty cargo space. And that's when the magic happens: we would take the calculations of fruits and vegetables likely to be harvested in a given farm, and cross with the idle space available in our vehicles. And based on the vehicles' actual or projected trip, the system will automatically assign pick up and deliveries of fruits and vegetables to soup kitchens located in the proximity of their planned route or seizing the negative bandwidth (i.e. the return trip that normally brings the vehicle empty). We would ask beneficiaries (3) to complete a survey about consumption and nutritional profiles in order to train our recommendation model. Our goal with them is to improve nutritional habits. Finally, based on the nutrition profile that we have built for every beneficiary (based on the supermarket receipts that they send us, crossed with the data from our deliveries to them and the consumption patterns of other beneficiaries located nearby) we can assign deliveries of fruits and vegetables to those soup kitchens who seem to need it the most.
If you have any questions