Group Data & AI·Alliance Bank
One platform for the Group's AI,
from use case to value.
Take a use case in, build it on Vertex AI over the Group lakehouse, have Model Risk Management validate it independently, deploy it in Google Cloud Malaysia, and reuse it across business units through the marketplace, with the value it returns tracked to the ringgit.
0
Assets in the catalogue
0
Business units and functions
0.0M
Calls / 30 days
0%
Customer data in-region
5.2M inference calls served from Google Cloud Malaysia in the last 30 days.
One platform
The whole AI lifecycle, on the Group lakehouse
Without a platform, a model moves between tools, teams and spreadsheets on its way to production, and what it returned is hard to show. The Alliance AI platform runs every stage in one place, so a use case moves from request to validated, in-region, measured production without changing hands.
- 01
Use cases in
Business units raise a need with an owner and a business case. Reuse is checked before anything is built.
Use-case pipeline
- 02
Build
Data scientists build on the Group lakehouse with governed, masked data products.
Vertex AI Workbench · Pipelines · BigQuery
- 03
Validate
Model Risk Management validates independently, by risk tier, before any business decision relies on it.
Second-line validation
- 04
Deploy in-region
Promoted through the model registry to endpoints in Google Cloud Malaysia, with versioning and rollback.
Vertex AI registry · endpoints
- 05
Monitor and reuse
Drift, quality and usage are watched in production, and every validated asset is listed for other BUs to reuse.
Model monitoring · Marketplace
- 06
Track value
Consumption is charged back in MYR and realised value is measured against the business case.
Chargeback · AI portfolio board
Data residency by construction
Customer data stays in Malaysia
Residency is not a clause appended to an outsourcing assessment. It is how the platform is built: the lakehouse, the training and the inference all run in Google Cloud Malaysia, and the paths that would take customer data out are closed.
- Production and customer data stay in Google Cloud Malaysia
- Singapore is for platform build, testing and non-sensitive pilot data only
- No customer data in prompts; Gemini is served on Vertex AI in-region
- Every access to customer data is purpose-bound, attributable and logged
The marketplace
Build it once, and let every business unit reuse it
The marketplace is the platform's shelf. Instead of each business unit commissioning its own churn model or rediscovering the same data, a validated asset is published once, with its owner, tier and terms, and reused across the Group.
Models
Scoring, forecasting, document, speech and GenAI models, each with a named owner, a risk tier and a validation record from Model Risk.
5.2M calls served in 30 days
Datasets
Lakehouse data products with Dataplex lineage, quality scores and a classification, so fitness for use is judged in seconds.
Classification and refresh cadence on every record
Agents
Agents that run real bank casework with declared tools, scoped data access and a named human at every consequential step.
11K runs in 30 days
Prompts
Reviewed, versioned Gemini prompt templates, so every business unit does not rediscover how to ask the same question well.
Fork into your own namespace
Demos
Working mini-applications built on catalogue assets: see a capability in a real journey before you commit to it.
Runnable in the browser
Sandbox
Vertex AI Workbench projects to evaluate, compare and retrain against masked lakehouse data, inside the region.
Drawn from your BU compute budget
Most reused
What the Group is actually using
Ranked by measured consumption rather than by whoever made the best case at steering committee. The same meter drives chargeback, so usage and cost are read off one record.
Models
1card-fraud-detectorAnomaly Detection · Financial Crime1.8Mcalls / 30d2payment-anomaly-monitorAnomaly Detection · Trade & Payments1.3Mcalls / 30d3transaction-narrative-nerLanguage AI · Financial Crime684Kcalls / 30d4next-best-offerScoring · Marketing312Kcalls / 30d5lead-priority-scorerScoring · Consumer Banking212Kcalls / 30dDatasets
1product-catalogueBigQuery table + JSON · 486 products and variants241downloads2branch-atm-networkBigQuery table · 1,284 rows214downloads3macro-indicatorsBigQuery table · 96K observations205downloads4casa-transactionsBigQuery table (partitioned) · 2.4B rows (36 months)186downloads5customer-360-martBigQuery table (partitioned) · 38.4M rows164downloadsAgents
1complaint-triage-agentCustomer Service Triage · Operations3.1Kruns / 30d2fraud-case-agentCustomer Service Triage · Financial Crime2.4Kruns / 30d3data-quality-agentData QA · Group-wide1.9Kruns / 30d4gov-drift-watchGovernance · Group-wide1.2Kruns / 30d5loan-doc-review-agentDocument Processing · Business Banking640runs / 30dBuild, validate, reuse
From notebook to Group asset, with the second line in the loop
Any team with a governed data product and a Vertex AI project can build here. Nothing reaches production without independent validation by Model Risk Management, and once live, what other units consume is charged back to them.
Build on Vertex AI
Workbench and pipelines over the Group lakehouse, with masked data products pre-mounted and experiments tracked.
Package and document
Model card, intended use, limitations, lineage and risk tier, captured once in the wizard rather than in a document.
Independent validation
Model Risk Management validates by tier: performance, fairness on credit decisions, explainability, and a named approver.
Live and reusable
Deployed to Google Cloud Malaysia and listed in the Group catalogue. The owner keeps accountability; other BUs get reuse.
And then it pays its way
Reuse is metered and charged back to the consuming business unit, in MYR
Every call, run and query is attributed to a cost centre. Each BU works inside a compute budget, and the showback tells Group Finance which assets earn their running cost and which do not. No BU has to raise an internal invoice to another.
The use-case pipeline
Reuse first, then build what is worth building
Every request enters the same pipeline. It is matched against what already exists, prioritised on value and data readiness, and tracked after go-live against the business case that justified it.
- 012 in stage
A use case is raised
A business unit states the decision it wants to improve, the owner accountable for the value, and the data it thinks it needs.
Use-case pipeline - 02
Reuse is checked first
The request is matched against the catalogue before a line is written. Often a validated model or data product already covers most of it.
Catalogue - 034 in stage
Prioritised on value
The business case, data readiness and risk tier are weighed side by side, and the AI Centre of Excellence sets the build order.
Pipeline - 043 in stage
Built and validated
Built on Vertex AI over the lakehouse, then independently validated by Model Risk Management before anyone relies on it.
Operating model - 053 in stage
Live, with value tracked
Once live, usage, drift and realised value are measured against the business case, telling the Group what to scale, fix or retire.
What is live The loop closes
A catalogue tells you what exists. The pipeline tells the Group what to build next, and the value record proves whether the last decision was right.
From SAS to the lakehouse
The move off SAS is already running on the platform
Retiring SAS is not a separate programme with its own tools. The rebuilt models go through the same build, validation and registry path as everything else, and their owners work on the same lakehouse as the rest of the Group.
0
SAS models rebuilt in Python
0
SAS users moved to the lakehouse
Rebuilt from SAS, now in the catalogue
Who publishes
Every business unit on the same shelf
A business unit's model sits beside a Group function's and an approved vendor's, judged on the same published metrics and cleared through the same validation gates.
Ready for the working screens?
77 assets with filters, validation status and usage, the catalogue practitioners work in every day.