Modules
Field vision
What it does
Takes an image or a field reading and marks what the operation should look at in this window. It does not replace an agronomist. It covers what one pair of eyes misses.
What you get
- A cut of the plot you sent
- What the model saw, with the model version
- What this reading left out
Data pipeline
What it does
Puts images, weather, machines, and handwritten notes into one history. Cloud Storage on intake. BigQuery for the tables.
What you get
- Versioned data, with a source
- A table ready for the next training run
- A trail of what came in and what was dropped
Models
What it does
The crop problem becomes an experiment. Train, score, and keep the version on Vertex AI.
What you get
- Model version and the assumptions behind it
- What the model does not cover
- A cutoff agreed with the operation
Inference
What it does
An endpoint the operation calls. The answer stays in that frame. Cloud Run or GKE when volume grows.
What you get
- A reply with the model version used
- A log of the question, for the next pass
- A clear miss when the data does not fit
Grower dashboard
What it does
The face of the product. Recommendation, history, and the next step. The person in the field does not open a Colab.
What you get
- A screen for the season or the plot you follow
- A recommendation in the operation’s words
- How to send the next batch
Architecture
The stack
Models train and serve on Vertex AI. History sits in BigQuery. Raw files go to Cloud Storage. The API runs on Cloud Run.
Vertex AI — train and serve
BigQuery — history
Cloud Storage — raw files
Cloud Run — API
GKE — when the load asks for it
IAM — who can see what
