How did you approach combining AI recommendations with real environmental data in Virdis?

[atomic-crater]

I recently explored Virdis and one part that stood out to me was the crop planning system, especially how it combines satellite imagery, soil data, weather analytics, land suitability scoring, and AI generated recommendations together.

A lot of projects use AI in isolation, but Virdis seems to integrate actual environmental and geospatial signals into the recommendation process, which feels much more practical and grounded.

I was curious about a few things:

• How do you structure the data before sending it into the AI workflow?
• How do you ensure the recommendations remain realistic instead of overly generic?
• Why did you choose a dual approach with both a local agronomy model and Gemini based planning?
• Did you face challenges balancing deterministic scoring systems with AI generated outputs?

Would genuinely love to hear more about the thinking and architecture behind this part of the platform.