Discussion cards

Eight decisions to settle in the meeting

For each: the question, the realistic options with trade-offs, what we currently know, what we still need to learn, and which apps + layers it touches. Where we have a working leaning, it's marked.

Decision 01

Cloud Infrastructure

GCP, AWS, Azure, or hybrid — and what drives the decision?

Bunge already uses GCP. The platform is cloud-native and must avoid on-prem dependencies. Multi-tenant from day one.

GCP-primaryWorking leaning

+ Pros

· Bunge alignment
· BigQuery + Vertex AI fit GeoAI
· Lower friction for Phase 1

− Cons

· Esri tooling more battle-tested on AWS
· Lock-in risk for non-Bunge tenants

AWS

+ Pros

· Strongest Esri partner ecosystem
· Mature multi-tenant patterns

− Cons

· Net-new for Bunge integration
· Data egress costs from GCP

Hybrid (GCP for Bunge, abstracted core)

+ Pros

· Tenant-aware infra from day one
· Future-proofs multi-customer

− Cons

· Higher engineering overhead
· Slower Phase 1

What we know

✓ Bunge runs on GCP
✓ ArcGIS Online is multi-cloud

Still to learn

? Bunge data residency requirements
? Cross-customer anonymization layer location

Affects:OriginationLogisticsProcessingTrade FlowSimulation

Decision 02

Salesforce Integration

What do we know about Bunge's Salesforce data model and the cadastre project underway?

Origination depends on Salesforce as a source of truth for accounts, interactions, and (in-flight) cadastre. Need to align rather than duplicate.

Embed Bunge AgriIntel web maps inside SalesforceWorking leaning

+ Pros

· No new UI for merchandisers
· Reuses Salesforce auth

− Cons

· UX limited by Salesforce shell

External Bunge AgriIntel app with bi-di sync

+ Pros

· Full UX control
· Easier cross-tenant

− Cons

· Two systems to learn
· Sync conflicts

What we know

✓ Bunge has a cadastre project in flight

Still to learn

? Current Salesforce object model for parcels / producers
? Ownership of the canonical cadastre
? Sync cadence + conflict rules

Affects:OriginationField App

Decision 03

Satellite Pipeline

Which providers first, and what does the processing architecture look like?

Sentinel is free but coarser. Maxar / Planet add resolution + revisit at cost. New constellations (SAR, hyperspectral) on the horizon.

Sentinel-first, Planet on demandWorking leaning

+ Pros

· Cost predictable
· Covers Phase 1 origination needs

− Cons

· Cloud cover limits revisit

Planet baseline + Maxar tasking

+ Pros

· Best revisit
· Operational-grade detail

− Cons

· High commercial cost
· Vendor concentration

What we know

✓ Origination + traceability + logistics all need imagery

Still to learn

? Tile processing in cloud-native or via Esri Image Server
? Storage tiering policy

Affects:OriginationLogisticsTraceability

Decision 04

AI Model Strategy

Build from scratch, leverage Esri GeoAI, use Vertex AI / SageMaker, or combine?

AI runs continuously, not on demand. Models must be reproducible across tenants.

Esri GeoAI for spatial, Vertex AI for everything elseWorking leaning

+ Pros

· Fast Phase 1
· Native ArcGIS integration

− Cons

· Esri model catalog is opinionated

Build proprietary models on Vertex AI

+ Pros

· Differentiation
· Full control

− Cons

· Time + cost
· Reinventing CV wheels

Cloud ML services only (no Esri GeoAI)

+ Pros

· Cloud-native
· Avoids Esri lock-in

− Cons

· Custom spatial pipelines

What we know

✓ GeoAI engine is a platform layer, not a per-app concern

Still to learn

? Model versioning + tenancy strategy
? Inference SLAs

Affects:OriginationProcessingTrade FlowSimulation

Decision 05

Mobile Platform

Native vs. cross-platform? Offline-first architecture for rural coverage?

Merchandisers operate in low-connectivity areas. Capture must work offline, sync later.

React Native + offline-first SQLiteWorking leaning

+ Pros

· Single codebase
· Strong offline story

− Cons

· Heavier native modules for maps

Native iOS + Android

+ Pros

· Best map + camera UX

− Cons

· 2x build cost

Esri Field Maps + custom shell

+ Pros

· Battle-tested offline maps

− Cons

· UX constrained by Esri app

What we know

✓ Field app feeds origination + cadastre

Still to learn

? Device fleet
? Conflict resolution rules on sync

Affects:Field AppOrigination

Decision 06

ArcGIS Deployment

ArcGIS Online, ArcGIS Enterprise, or hybrid?

Esri is the GIS backbone. Most users will never open a GIS app.

ArcGIS Online (managed)Working leaning

+ Pros

· No infra
· Fast Phase 1

− Cons

· Less control over data residency
· Per-credit cost model

ArcGIS Enterprise on GCP

+ Pros

· Full control
· Data stays in tenant cloud

− Cons

· Heavy ops
· License cost

Hybrid (Online for collaboration, Enterprise for heavy compute)

+ Pros

· Cost-optimized
· Flexible

− Cons

· Two surfaces to manage

What we know

✓ ArcGIS is non-negotiable as the GIS backbone

Still to learn

? Named-user vs. app-login licensing
? Imagery compute footprint

Affects:OriginationField App

Decision 07

Data Pipeline

Real-time vs. batch? Event-driven? How do we handle multi-source ingestion at scale?

Sources include Salesforce, SAP, satellites, AIS, USDA, mobile capture — different cadences, different SLAs.

Event-driven (Pub/Sub) + batch for imageryWorking leaning

+ Pros

· Right tool per cadence
· Scales horizontally

− Cons

· Two pipelines to operate

All batch (nightly)

+ Pros

· Simple

− Cons

· Breaks vessel tracking + field sync

All streaming

+ Pros

· One model

− Cons

· Overkill for imagery + cost

What we know

✓ Vessel tracking + field capture need near-real-time

Still to learn

? Bunge SAP extraction pattern
? CDC vs. polling for Salesforce

Affects:LogisticsTrade FlowTraceabilityField App

Decision 08

Cross-cutting — Multi-tenancy + Data Sovereignty

How do we isolate customer data while still surfacing cross-customer insights?

Architecture principle: customer data isolated, cross-customer insights only from anonymized patterns.

Per-tenant project + shared anonymization layerWorking leaning

+ Pros

· Hard isolation
· Clean audit story

− Cons

· Anonymization layer is non-trivial

Single project with row-level tenancy

+ Pros

· Cheaper
· Faster cross-cuts

− Cons

· Higher leak risk

What we know

✓ Bunge is tenant #1 but not the only one

Still to learn

? Legal definition of anonymization per geography

Affects:OriginationLogisticsProcessingTraceabilityTrade FlowSimulationField App