Architecture
How the platform is built, and what changes on the path to production
Channels
Where the customer actually is. Nothing here is a new banking app the customer has to be persuaded to install.
Experience layer
Role-aware interfaces. Six personas see six different versions of the same underlying data.
AI orchestration
The layer that makes it an agent rather than a chat window: it selects tools, records what it touched, and stops at the human gate.
Model layer
All bank-resident. No public internet egress, no customer data leaving the perimeter.
Integration layer
One interface per capability. The vendor behind it is a configuration detail, which is what makes the platform portable.
Core banking
Unchanged. The platform reads and writes through the adapter and replaces nothing the bank already runs.
End-to-end data flow
- 1Customer sends a document over WhatsApp
- 2Messaging adapter normalises and stores it
- 3Document AI classifies and extracts fields with confidence
- 4Low-confidence fields route to a human queue
- 5Risk model re-scores using the new evidence
- 6Explainability engine attaches factors to sources
- 7Analyst reviews; approver records the decision
- 8Adapter writes the loan record to core banking
- 9Notification goes back out on the customer's channel
- 10Every step above writes one audit row
What is simulated in this demo
- All customer, application and document data — fictional
- Core banking responses — generated locally by mock adapters
- Message delivery states — no channel provider is contacted
- Model inference — deterministic routing, not a live model call
- Policy documents — written for this demo, not regulatory text
Adapters implemented
- Temenos T24T24AdapterSimulated
- Oracle FLEXCUBEFlexcubeAdapterSimulated
- Futeric Core BankingFutericCoreAdapterSimulated
The path to production
What stays as built
- Domain model and types
- Experience layer and RBAC
- Orchestration logic and approval gates
- Execution trace and audit writer
- Design system and bilingual support
What gets replaced
- Mock adapters with production adapters
- Seed data with real datastores
- Deterministic routing with real model inference
- Local policy file with a vector store
- In-memory state with a persistence layer
What must be added
- Authentication and enterprise identity federation
- Encryption in transit and at rest
- Monitoring, alerting and tracing
- Model versioning and rollback
- Penetration testing and security review