AI and the Potato Industry

What is ETL?

ETL stands for Extract, Transform, and Load. It’s a standard way to move data from where it’s collected into a form that’s useful for reporting, dashboards, or AI. Once the data is cleaned and structured, you can plug in both machine learning (ML) models and generative AI (GenAI) to get numbers, predictions, and human-readable text.

• Extract: Pull data from field sensors, weather APIs, spreadsheets, storage logs, or equipment.
• Transform: Clean, standardize, and combine data (e.g. fix units, fill gaps, join field + weather).
• Load: Write the result into a database, data lake, or analytics tool so people and AI can use it.

AI fits into this pipeline after (or during) the transform step: you feed the prepared data into ML models for predictions and into GenAI for summaries and alerts. Then you load both the raw outputs and the generated text to dashboards, apps, and SMS.

ML vs GenAI in the Pipeline

Machine learning (ML) uses patterns in your data to produce structured outputs: numbers (e.g. predicted yield in t/acre), categories (e.g. high/medium/low risk), or yes/no flags. ML models are trained on historical data and run on the same kind of structured data from your ETL pipeline. They don’t write sentences—they output columns you can chart, filter, and send to a dashboard or app.

Generative AI (GenAI) produces text (or images). In a pipeline, GenAI takes the same data—and often the ML outputs—and turns them into plain language: weekly summaries (“Field F102 is trending below target; consider extra scouting”), SMS alerts, answers to questions like “Why is field F101 predicted higher than last year?”, or short reports for management. GenAI doesn’t replace ML; it makes ML results easier to read and act on.

ML examples (potato)

• Yield prediction: Regression model trained on field + weather + history → predicted t/acre per field.
• Disease/risk classification: Model trained on weather + soil + past outbreaks → “low / medium / high” risk per field or week.
• Grade or quality: Model trained on storage temp, humidity, variety → predicted grade or “process vs table” suitability.

GenAI examples (potato)

• Weekly summary: “This week F101 and F103 are on track. F102 is 0.4 t/acre below target; wet conditions in the north corner may be a factor.”
• SMS alert: “Alert: Field F102 yield forecast dropped. Consider scouting. Details in app.”
• Q&A over your data: User asks “Which fields need attention?” → GenAI reads ML outputs and answers in plain language.

Potato Example: From Field Data to ML and GenAI

Below is a single pipeline: data is extracted from several sources, transformed into one table, then fed to an ML model for yield and risk, and to GenAI for summaries and alerts. The results are loaded to a database and then surfaced on a dashboard, in an app, and via SMS.

Step 1: Extract (data sources)

Data is pulled from sources typical in potato operations:

• Field records: variety, planting date, field ID, acreage.
• Weather: temperature, rainfall, growing degree days (from a weather API or station).
• Soil/sensors: soil moisture, pH, or other sensor readings if available.
• Historical yields: tonnes per acre or per field from past seasons.

Step 2: Transform (clean and join)

Raw data is cleaned and combined into one table per field-season: same units, same date ranges, and no missing key values. Example shape after transform:

Step 3: Plug in ML and GenAI

The transformed table is the input to an ML model. The model adds predicted yield and risk level per field. Those outputs (plus the underlying data) are then passed to GenAI to produce short summaries and alert text.

Example table after ML (two new columns):

GenAI takes this table (and optional context like “weekly update”) and generates, for example: “F101 and F103 on track. F102 below target (16.9 t/acre); medium risk—consider scouting.” That text is stored and also sent as SMS or in-app copy. The pipeline runs on a schedule (e.g. weekly) so predictions and messages stay current.

Step 4: Load — Where results go

The final dataset—including ML predictions and GenAI text—is loaded into a database or data store. From there it powers a web dashboard, a mobile app, and SMS (or email). The next section shows example outputs for each.

Output to Dashboard, App, and SMS

The same pipeline can feed multiple surfaces. Below are mock examples of how the ML and GenAI outputs could look on a dashboard, in an app, and in an SMS.

Dashboard (web)

A simple web dashboard might show key numbers and a chart. Tiles can come from the loaded ML results; the “Weekly summary” text is from GenAI.

App (mobile)

The same metrics and summary can appear in a mobile app. Alerts can be push notifications; the content can be the same GenAI summary or a shorter line.

SMS

SMS messages are kept short. GenAI can generate one or two lines from the ML results, e.g. for alerts or a daily/weekly digest.

Why This Pipeline View Helps

Keeping ETL, ML, and GenAI in one pipeline makes it easier to add new data sources, new models, and new ways to consume the results. Good AI depends on good data; ETL forces you to define what you extract and how you transform it before adding models.

• Data first: One clean dataset feeds both ML and GenAI, so predictions and messages stay consistent.
• Reusable: The same pipeline can feed multiple ML use cases (yield, risk, quality) and multiple GenAI outputs (summary, SMS, Q&A).
• One load, many surfaces: Load results once to a database; the dashboard, app, and SMS all read from the same place.