# Building a customer data platform on a sovereign lakehouse: the AI360 architecture, step by step

> A lakehouse-native CDP builds the customer profile where the data already lives. In DataNature, Connect and StreamHouse land customer data in Bronze, dbt resolves identities in Silver, AI360 Studio builds Gold profiles, features, segments and scores, and reverse ETL, Lakebase and Voir activate them. DataNature Atlas enforces consent, purpose and masking throughout, all inside your jurisdiction.

Source: https://datanature.ai/blog/customer-data-platform-architecture-ai360-lakehouse
Publisher: Dlytica Inc. (DataNature is a Dlytica company)
Updated: 2026-11-05

The composable CDP idea is simple: bring CDP functions to the data instead of copying the data into a CDP. Doing it well is an engineering job with clear stages. This guide walks through how AI360 Studio implements a customer data platform on DataNature, with the design decisions at each step.

## What does the architecture look like end to end?

| Layer | Component | Output |
|---|---|---|
| Sources | Core banking or billing, CRM, apps and web, contact centre, card processor, documents | Raw customer records and events |
| Ingest | Connect (batch), StreamHouse (real time), DocuContext (documents) | Bronze Iceberg tables per source |
| Unify | dbt identity-resolution models in Silver; knowledge graph | One `customer_key` per person, linked identifiers, households |
| Profile | Gold customer 360 and feature store | Attributes, holdings, behaviour, consent state |
| Decide | AI360 Studio models, segments and decisioning; AI Twins built with Voir | Scores, segments, next-best actions, staff briefings |
| Activate | Reverse ETL, Lakebase profile API, Voir agents | Audiences in tools, profiles in apps, briefings for staff |
| Govern | [DataNature Atlas](/platform/atlas) | Consent and purpose policies, masks, lineage, audit |

## How does identity resolution work?

Identity resolution runs as versioned dbt models in Silver, so the rules are reviewable code, not a black box:

1. **Standardize** identifiers: normalise phone numbers, emails, names, addresses and national ID formats.
2. **Deterministic matching** on strong keys: customer number across products, verified email, verified phone, national ID where permitted.
3. **Probabilistic matching** for the rest: weighted similarity on name, date of birth, address and device, with a threshold and a review queue for borderline pairs.
4. **Clustering** matched records into one `customer_key`, with a crosswalk table from every source ID to the key.
5. **Survivorship rules** choose the best value for each attribute, such as the most recent verified address.
6. **Households and relationships** are added as edges in the knowledge graph: joint accounts, guarantors, shared addresses.

Keep the crosswalk and every match decision, so you can explain why two records were merged and unmerge them if needed.

## What goes into the profile?

The Gold profile combines four kinds of data:

| Kind | Examples |
|---|---|
| Identity and contact | Resolved name, verified contacts, preferred channel, language |
| Relationship | Products held, balances or plan, tenure, branch or region, household |
| Behaviour | Transactions, app and web activity, service contacts, complaints |
| Permissions | Consent by purpose and channel, opt-outs, data-sharing preferences |

Behavioural aggregates (spend by category over 90 days, logins per week, days since last contact) go into the feature store, so marketing, risk and AI models use the same definitions.

## How are segments and scores built?

AI360 Studio works on the Gold profile and features:

- **Segments** defined by business rules or clustering, stored as governed tables with owners and refresh schedules.
- **Predictive scores**: churn, propensity per product, lifetime value, credit or fraud risk where relevant, each versioned in the model registry with the data snapshot it was trained on.
- **Decisioning**: rules plus scores produce the next-best action per customer, respecting eligibility and consent.
- **AI Twins**, built with Voir: persistent, domain-trained replicas of key customers, relationships and analysts that reason over the AI360 profile and advise staff.

## How does activation work, batch and real time?

- **Reverse ETL** pushes audiences and attributes to campaign, CRM and service tools on a schedule.
- **Real-time profiles**: StreamHouse updates short-window features within seconds, and Lakebase serves profile lookups to apps and channels with millisecond reads.
- **People**: Voir agents brief relationship managers, branch and contact-centre staff with the profile and next-best action, answering from governed cubes.
- **Feedback**: responses and outcomes return to the lakehouse to measure lift and retrain models.

## How are consent and privacy enforced?

Consent is part of the profile, and [DataNature Atlas](/platform/atlas) enforces it:

- Activation jobs check consent by purpose and channel; audiences that would use data outside the consented purpose are blocked.
- Personal data is masked by role: analysts see hashed identifiers that still join; only exempt roles see clear values.
- Every activation, export and agent answer is logged with lineage back to source.
- The whole CDP, including AI inference, runs in your chosen jurisdiction.

## What is a sensible rollout order?

1. Land two or three core sources and build identity resolution for the main customer key.
2. Publish a governed customer 360 dashboard and check it with the business.
3. Add consent and the feature store.
4. Ship one scored use case, such as churn or next-best product, with reverse ETL to one channel.
5. Add real-time events through StreamHouse for in-session use cases.
6. Add Voir briefings for staff once the numbers are trusted.

## Frequently asked questions

### Do we need a separate CDP database?

No. Profiles live in Gold Iceberg tables, features in the feature store, and low-latency lookups in Lakebase, all on the same governed platform.

### Can identity resolution run in real time?

New events attach to an existing `customer_key` in real time through deterministic matches; probabilistic matching and re-clustering run in batch.

### How long does a first version take?

Typically one wave of a foundation-first programme: a few months for core sources, identity resolution, the 360 dashboard and a first scored use case.

## Sources

- [CDP.com: what is a CDP, including identity resolution](https://www.cdp.com/basics/what-is-a-cdp/)
- [DataNature: dbt on Spark and Trino with Iceberg](/blog/dbt-spark-trino-iceberg-medallion-pipelines)
- [DataNature: StreamHouse, the real-time hot path](/blog/streamhouse-real-time-hot-path-lakehouse)
- [DataNature solution: customer 360 with reverse ETL](/solutions/customer-360-reverse-etl)

## Where do website, app and voice channels fit in the architecture?

- **Behavioural collection** sits beside Connect and StreamHouse in the ingest layer: AI360 event SDKs capture website, app and server events on your own domain, run in-flight filters and masks, and stitch anonymous IDs to known users before identity resolution in Silver. See [website and app event tracking](/blog/website-app-event-tracking-sdk-cdp-identity-stitching).
- **Voice** is an activation channel: AI360 sends approved per-customer variables to a voice agent platform such as ElevenLabs Agents, and call outcomes, answers and opt-outs return by webhook into the profile. See [AI voice calls from your CDP](/blog/ai-voice-outbound-calls-cdp-hyper-personalization).

## How does DataNature Atlas keep this secure and compliant?

Governance is part of the architecture, not a separate tool bolted on later. [DataNature Atlas](/platform/atlas) is the platform's governance module:

- **Knows your data**: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
- **Controls access**: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
- **Governs AI**: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
- **Proves it**: lineage from source to metric to decision, and an audit log of who or what accessed which data.
- **Stays local**: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.

These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read [DataNature Atlas in depth](/blog/datanature-atlas-governance-rbac-abac-pii-masking) or see [security and compliance](/security).
