Farmly

Our Vision

Kirchenfeldrobotics was invited to the Swiss{ai}weeks Hackathon in Bern, where we aimed to design a mobile application that would improve the workflow of local farmers. In Switzerland, many farmers sell goods in small local farm shops, where people can buy cheese, sausages, pumpkins, and more. These goods have to be delivered to the shops, and this is usually done by the farmers themselves. This process is very inefficient.

We tried to solve this problem with Farmly, developed in just 24 hours. Our idea was to use farmers who are already heading to their shop. They could pick up merchandise from other farmers on their route—provided the destination of that merchandise matched their direction. This is often the case, since many farmers live in similar areas and sell their goods in similar regions. Farmly would essentially create a shared network for local farm-goods deliveries.

Realization

We tried to use the Apertus 70B LLM for our application, since it was a newly released Swiss AI model and we were excited to experiment with it. We also appreciated that Swisscom sponsored API keys for this model.

Just like at our first hackathon, we aimed to build a scalable solution that could demonstrate what a fully developed app might look like. We built the app using the Flutter SDK and used Google Firebase for the database.

The main idea of the app was to allow users to chat with an AI chatbot that could detect whether they wanted to send a package or whether they were driving somewhere and able to deliver one. The Apertus 70B API was intended to serve as the chatbot’s AI engine. The route a farmer planned to take was calculated using the OpenRouteService API. We this API, and some additional libraries, we were also able to calculate deployed merchandise on this route.

The app included several screens that allowed farmers to keep track of their goods and the routes they planned to travel.

Issues we encountered

For this idea to work, the AI had to be agentic. It needed to recognize specific scenarios and trigger the correct function calls. Some sources suggested that Apertus had these capabilities. It turned out that it did not. As a result, our entire implementation was doomed to fail. However, all other parts of the app worked perfectly: the only issue was that the AI could not trigger the functions e.g. required to handle the deployment of merchandise.

Our Achievements

Even if our app faced a fundamental technical obstacle, we remain convinced that its fully developed version could significantly improve the daily workflow of local farmers. The concept itself was strong, and the prototype already demonstrated how such a system could streamline local logistics and support small-scale producers. Ultimately, this is what a Hackathon is truly about: experimenting, learning fast, and building solutions that have the potential to create real-world impact. Even if every piece isn’t perfect yet.