A farm that’s run by talking in Catalan

The challenge

Mas Oller is a family farm in the Baix Penedès with a vineyard, a vegetable garden, fruit trees and plants, shared by several families. Nobody wanted yet another app. They wanted to ask questions, jot things down and get alerts on their phones, in their own language and without having to learn anything.

What we built

A Hermes agent with its own Discord server and a channel for each thing: the vineyard, the vegetable garden, the plants, expenses and administration.

• Real weather data. The farm’s Ecowitt station feeds into Home Assistant and is stored in InfluxDB, with 188 days of recovered history and 736,005 data points. Every night four agronomic indicators are calculated: infection risk for the vines, stress and pests in the vegetable garden, plant hydration and an alert for the next 24 hours.
• Diagnosis by photo. A general vision model and one specialized in crop pests give two opinions; the agent synthesizes them with a confidence level and asks for more photos if needed.
• Field log. Irrigation, treatments with product and dose, fertilizing and maintenance, as free text or as a voice note in Catalan or Spanish. Each entry records who did it.
• Plant register with stable identities and autocomplete, to follow each plant’s progress photo by photo.
• Shared expenses between families, with automatic settlement of who pays whom.

How it’s used

Emilio, who looks after the vineyard, sends a voice note along the lines of “today I treated the five rows with sulfur, two kilos” and it gets logged with the date, product and author. If he uploads a photo of a leaf, he gets a reasoned diagnosis and the questions still needed to narrow it down.

What I learned

• The interface is the conversation. Short commands, voice and the user’s own language.
• Data gets precomputed. Calculating four indicators every night is far more reliable than asking the model to go through thousands of data points.
• You have to check against reality. I validated the weather data point by point against the station’s physical display and found three errors that a model would never have caught.
• The agent must not touch the infrastructure from the family channels. I blocked it by design.