Calibration
Does the model’s confidence match what actually happens? We compare predicted yield, marketable life and likely diagnoses with measured results.
Home / Research
Research tested where farming happens.
Our research asks one question: can agricultural advice be trusted when it matters? We study it with farmers, extension workers and researchers, and we judge it by what happens in the field.
Research that asks the right questions.
Five questions shape our work. Each one is studied in the field, not only in the lab.
Human oversight
When should advice defer to a person? We study how uncertainty is detected, explained and routed to agronomists and extension workers.
Farm memory
How much does context improve advice? We test what difference crop, planting date, growth stage and recent work make to a recommendation.
Low-connectivity delivery
What works on a low-cost phone on an unstable network? We study offline capture, retry without duplicates, and simple interactions.
Local knowledge and language
How well does advice fit local crops and practice? Our work considers cassava, tomato and maize, alongside the language and practical context of the farmer.
Go deeper
Five questions, each studied in the field.
Choose a question to see how we study it and what we look at.
Calibration
Research question
Does the model’s confidence match what actually happens?
We compare predicted yield, marketable life and likely diagnoses with measured results.
- We look at
-
- Predicted values against what was measured later
- Stated confidence alongside accuracy
- Cases where a confident answer was wrong
Human oversight
Research question
When should advice defer to a person?
We study how uncertainty is detected, explained and routed to agronomists and extension workers.
- We look at
-
- How thin evidence is detected
- How the uncertainty is explained to the farmer
- Where the case is routed, and how quickly
Farm memory
Research question
How much does context improve advice?
We test what difference crop, planting date, growth stage and recent work make to a recommendation.
- We look at
-
- The same question with and without farm context
- Crop, planting date and growth stage
- Recent work and observations
Low-connectivity delivery
Research question
What works on a low-cost phone on an unstable network?
We study offline capture, retry without duplicates, and simple interactions.
- We look at
-
- Work captured offline stays intact
- Retries that do not create duplicates
- Interactions that stay simple on a basic phone
Local knowledge and language
Research question
How well does advice fit local crops and practice?
Our work considers cassava, tomato and maize, alongside the language and practical context of the farmer.
- We look at
-
- Cassava, tomato and maize
- The language a farmer uses
- Local practice and conditions
How we evaluate.
Every study follows the same loop, so results from different crops and regions can be compared.
-
01
Define the decision
Start from a real decision a farmer or field officer has to make, such as whether to treat a crop, when to harvest or how to store a batch.
-
02
Record the context
Capture the farm, crop, season and observations that the decision depends on, so every result can be traced to its inputs.
-
03
Compare with expert judgement
Agronomists and extension workers review the same case independently, and their judgement is the reference for quality.
-
04
Follow the outcome
Compare what was predicted with what was measured later in the season: yield, condition at harvest, marketable life.
-
05
Report what we learn
Share results with collaborators, including the cases where the model was wrong, and let the evidence decide what is built next.
One loop, every study
What we learn decides what we build next.
Results from one study feed the next question, so different crops and regions can be compared.
Field conditions in three regions.
Our pilots span savanna, coastal and delta conditions, so a result is tested against more than one environment.
- Regional focus
- Federal Capital Territory, Lagos and Rivers
- Crops
- Cassava, tomato and maize. More are added only after their data and model behaviour are reviewed.
Three kinds of ground
A result is tested against more than one environment.
Our pilots in the Federal Capital Territory, Lagos and Rivers span savanna, coastal and delta conditions.
How we hold ourselves to account.
Uncertainty is reported, not hidden
We report calibration and stated confidence alongside accuracy, because a confident wrong answer costs a farmer more than an honest “not sure”.
Failures are findings
Cases where the model was wrong are recorded and studied. They guide what we change and which crops or features we hold back.
Limits are part of the result
Openfarms Lab does not guarantee yield, replace a qualified agronomist or certify produce as safe to eat, and our evaluations test those limits.
People’s data is protected
Farm and personal data are handled under our privacy notice and shared with research collaborators only as agreed for each study.
Research with us.
We work with universities, institutes and practitioners. If you would like to test our model independently, start with the benchmark page.
Evaluate our model independently.
We invite researchers and institutions to test the AI Extension Worker against their own methods.