AI Model & Decision Center

Registry · explainability · decision intelligence engine

Prototype model metrics — demonstration only.Models, metrics and lineage are simulated on synthetic data to illustrate governance — not validated production figures.

Registered models

13 business domains

In production

7 pilot · 2 in validation/dev

Predictions / day

210k

real-time + near real-time + batch

Drift on watch

RUL · OER · DIS · CPO

Human-approved actions

100%

no autonomous execution

Model registry

17 of 17 models · click a row for model card, lineage and explainability

Model	Domain	Algorithm	Version	Key metric	Stage	Inference	Drift	Predictions / day
Yield Prediction MDL-YLD-041	Plantation	XGBoost (gradient-boosted trees)	v4.1.0	MAE 1.42 t/ha (R² 0.86)	Production	Batch	Stable	50
Asset Failure Prediction MDL-AFP-032	Maintenance	Gradient Boosting + vibration spectral features	v3.2.1	AUC 0.91 (7-day horizon)	Production	Real-time	Stable	144,000
Remaining Useful Life MDL-RUL-011	Maintenance	Survival model (Weibull AFT)	v1.1.0	C-index 0.78	Pilot	Near real-time	Watch	2,400
OER Prediction & Optimization MDL-OER-024	Mill	LightGBM + constrained optimizer	v2.4.0	MAE 0.18 pp OER	Production	Near real-time	Watch	96
FFB Quality Grading (Vision) MDL-FFB-CV-013	Computer Vision	YOLO-style detector + ripeness classifier	v1.3.2	mAP@0.5 / F1 0.87 / 0.84	Pilot	Real-time	Stable	9,200
Inventory Demand Forecast MDL-INV-018	Supply Chain	Hierarchical time series (ETS + gradient boosting)	v1.8.0	MAPE 11.4%	Production	Batch	Stable	200
Spare Parts Prediction MDL-SPP-012	Supply Chain	Rules + failure-probability linkage	v1.2.0	Precision@10 0.82	Pilot	Near real-time	Stable	600
Disease & Pest Risk MDL-DIS-015	Agronomy	Random forest classifier + spatial lag	v1.5.0	AUC 0.83	Pilot	Batch	Watch	50
Fuel Anomaly Detection MDL-FUEL-013	Logistics	Isolation forest on trip features	v1.3.0	Precision 0.79	Production	Near real-time	Stable	1,100
Weighbridge Anomaly Detection MDL-WB-021	Governance	Autoencoder + rule ensemble	v2.1.0	Review hit rate 41% confirmed issues	Production	Near real-time	Stable	590
Harvest Forecasting & Scheduling MDL-HVF-020	Harvest	Prophet-style TS + MILP scheduler	v2.0.1	MAPE 7.9%	Production	Batch	Stable	50
Harvest-to-Mill Dispatch Optimizer MDL-RTE-010	Logistics	Vehicle routing (OR-Tools style) + queue simulation	v1.0.3	Queue time ↓ −38% (simulation)	Pilot	Near real-time	Stable	24
Fire Risk MDL-FIRE-011	Sustainability	Logistic GAM	v1.1.0	AUC 0.88	Production	Batch	Stable	50
Methane / Biogas Yield MDL-CH4-010	Sustainability	Mass-balance + gradient boosting	v1.0.0	MAPE 8.6%	Validation	Batch	Stable	2
CPO Price Scenario MDL-CPO-012	Commercial	Bayesian structural time series	v1.2.0	MAPE (30d) 5.8%	Pilot	Batch	Watch	1
Safety Vision (PPE / Proximity) MDL-SAFE-014	Safety	Object detection + pose + zone rules	v1.4.0	F1 0.86	Pilot	Real-time	Stable	52,000
Seed Performance Prediction MDL-SEED-010	R&D	Mixed model (G×E) + gradient boosting	v1.0.2	R² 0.74	Development	Batch	Stable	1

Explainable AI

Why a model predicted what it did — local (single prediction) and global (whole model) SHAP-style attributions

Local explanation · Block B17 · 12-month forecast

Base 27.8 t/ha (portfolio mean, same age class) → 24.4 t/ha (target 28.0)

Contribution to prediction (t/ha)

Rainfall deficit (3-mo lag)

−1.8

Late fertilizer application

−1.1

Palm maturity (15 yrs)

+0.9

Harvest interval 13 d

−0.7

Ganoderma pressure

−0.4

Soil moisture

−0.3

◀ lowers predictionraises ▶

Global feature importance (mean |SHAP|, normalised)

Rainfall (lagged 3/6/12 m)

0.28

Palm age / maturity

0.22

Fertilizer timing & dose

0.16

Harvesting interval

0.12

NDVI trend

0.09

Soil moisture

0.07

Pest & disease incidence

0.06

Plain-language explanation

Confidence88%

Weather explains the largest share of B17's gap, but two of the top four drivers — fertilizer timing and harvest interval — are controllable. Re-timing the K round and restoring a 10-day round would recover ≈ 1.8 t/ha.

Prototype attributions on synthetic data — shown to illustrate explainability standards.

Decision Intelligence Engine

How data becomes an approved action — and how outcomes flow back to improve the models

Counts are live from this session's Action Center — approve or reject an action and the loop updates.No recommendation is executed without a named human approver.