AI Model & Decision Center
Conceptual target architecture — from operational data to approved decisions
Layered data & AI architecture
Conceptual reference architecture — vendor-neutral
1Data Sources
ERP
Plantation Mgmt System
Mill System (MES)
SCADA / PLC
IoT Sensors
CMMS / EAM
GPS / Telematics
Weighbridge
Warehouse
Procurement
HR
Finance
Weather API
Satellite
Drone
CCTV
Laboratory
Manual field input
18 sources · ~53 M records/day
2Integration Layer
API gateway
Event streaming
Batch ETL
Change data capture
File & image ingestion
SCADA/IoT stream < 1 s
3Data Platform
Operational data store
Data lakehouse
Time-series DB
Geospatial DB
Document store
Bronze → silver → gold
4AI / ML Layer
Feature store
Model training
Model registry
Inference services
Model monitoring
17 models · MLOps
5Decision Intelligence
Business rules
Prediction
Optimization
Recommendations
Human approval gate
Explainable · auditable
6Applications
Executive dashboard
Mobile field app
Alerts & notifications
AI Copilot
Workflow & approvals
Web · mobile · chat
Security & access
Role-based access, estate/mill row-level security, encryption in transit & at rest
Data governance
Catalog, lineage from source to recommendation, data-quality SLAs per source owner
Feedback loop
Approvals, rejections and outcomes written back to the feature store for retraining
AI inference architecture
Three serving patterns, chosen by how fast the decision has to be made
REAL-TIME < 1 s end-to-endST-04 vibration crosses the bearing-defect signature → alert to Mill A maintenance in under a second.
NEAR REAL-TIME 1 – 5 minWeighbridge ticket WB-260928-0187 posted → features vs truck master & GPS → anomaly score 0.91 → routed for review.
BATCH nightly / monthlyMonthly yield forecast for 50 blocks — B17 flagged 24.4 vs 28.0 t/ha with SHAP drivers for the agronomist.