Four layers of knowledge. No black box.
The engine is built in four layers: an agronomic reference base, statistical and machine learning models, expert validation and proprietary season data. Today the Season Planner runs on the first layer, documented agronomic rules (FAO, MINADER) applied to sourced data. The models will be trained on the field data of the 2027 pilot, and each recorded season will make them more accurate.
From seven data domains to a decision on one plot.
Select any component to see its role, its technology and its release. Today the planner runs in your browser, and on our server for the PDF reports; the data and model pipelines of the coming releases will be written in Python.
Farmers · investors · cooperatives · lenders · institutions
Season Plan
Release · Available todayFeasible area, calendar, budget, water plan, risks, best window and confidence, computed in the browser and on the server for PDF reports.
Why the plans can be trusted before the models exist.
| Layer | Role | Source of reliability | Release |
|---|---|---|---|
| Agronomic reference base | Cycle lengths, water needs, temperature thresholds, crop management, standard costs per crop and zone. | Established research (FAO and national institutes), validated by agronomists. | Available today |
| Statistical and machine learning models | Climate risk, pest and disease risk, yield, harvest price, best window. | To be trained and tested on historical and pilot data, with measured error. | MVP 2027 |
| Expert validation | Review of cases the models cover poorly, local calibration. | Agronomists who know each zone. | Version 2 |
| Proprietary season data | Actual costs, yields, prices and incidents recorded by users. | Local, recent and unique to Djalumo Agro. | Grows each season |
Five questions, five models, five success metrics.
| Model | Question | Method | Training data | Success metric |
|---|---|---|---|---|
| Climate risk of a planting window | How likely are excess rain, heat at flowering or drought for this date and place? | Classification: logistic baseline, gradient boosting | 40 years of daily weather; agronomic thresholds by crop stage | Recall on adverse seasons, ROC AUC |
| Pest and disease risk | How favourable will conditions be for late blight, whiteflies or caterpillars? | Agronomic risk index, then supervised classification | Weather, crop stage, incidents reported by users | Precision and recall on reported outbreaks |
| Expected yield | What range of yield can this season deliver? | Regression: regularised linear baseline, random forest, XGBoost, with prediction intervals | Yield series with climate, then season logs at plot level | MAE and RMSE against a historical baseline |
| Harvest price | What price is likely in the harvest weeks, in which market? | Seasonal time series models and gradient boosting | Market price series and field price surveys | Mean absolute percentage error |
| Best crop and window | Which option offers the best margin for an acceptable risk within this budget? | Constrained optimisation over the model outputs | Outputs of the models above and the cost base | Margin at risk, user validation |
No prediction without its drivers and its uncertainty.
Feature attributions (SHAP) explain each estimate, and prediction intervals replace single numbers. This transparency builds trust with farmers and is a prerequisite for lenders and insurers.
High risk
In most of the last 40 years, heavy rain fell during flowering for this planting date.
Illustrative layout of an explanation. Directions only; no model output is shown.
Plan. Analyze. Construct. Execute.
Model development follows the PACE framework of the Google Advanced Data Analytics Professional Certificate held by the founder. Every model is compared with a simple baseline, validated in chronological order on years it has never seen, and audited for its errors by zone, crop and season before release.
Plan
Define the decision, the users, the success metric and the ethical limits.
Analyze
Audit sources, clean data, explore patterns by zone, crop and season.
Construct
Build baselines and models, validate on unseen years, explain with SHAP.
Execute
Release behind a gate, monitor errors, retrain as seasons are recorded.
Principles that hold under pressure.
- Predictions support decisions and never replace them.
- Uncertainty is always shown.
- No score is used automatically to exclude a farmer from credit or insurance.
- Underrepresented zones and profiles are monitored explicitly.
- Farmers can see, correct and export their own data.
Bring Djalumo Agro to your farms, portfolio or programme.
Pilot participation in Foumbot from March 2027, data partnerships, investment and institutional collaboration.