Data modeling

Bring AI to the modeling exercise

Agents help you reason over the business, profile the sources and propose the data model. The result is a data model that is machine readable, with its lineage and logic, and that engineers can query from the tools they already use.

Book a working session How the context store works

Where modeling stops today

The data model ends at the diagram

Classic modeling tools such as Erwin and PowerDesigner generate the DDL for the tables and stop there. The transformation code is written beside the data model by another team, the two drift apart, and when an AI initiative asks for shared context, the diagram cannot answer. The model owner ends up defending definitions nobody can execute.

A man and a woman discussing a diagram at a whiteboard
Modeling with an agent

The agent proposes, you decide, the graph remembers

Modeling happens in a session you can read, in Git and in shared chats, where people and agents review each other's proposals. The context carries from one piece of work to the next.

Reason over the business model

Asks clarifying questions, identifies concepts and keys, flags ambiguities such as account versus contract versus order, and writes the conceptual model into the context store.
Reason over the business model

Profile the source, map it to the data model

Profiling turns a business key from an assumption into a decision you can defend. The agent then proposes how each source maps to the data model.
Profile the source, map it to the data model

Review every proposal

The proposed structure is reviewed in Git or a shared session before anything is generated. People and agents review each other's work.
Review every proposal

The data model becomes a graph

Business context, implementation context and data product context in one versioned graph, open to any agent over MCP.
The data model becomes a graph
Jev, the Data Vault modeler: two sources and the business model become one Data Vault model, and you review only what is uncertain
Pick your approach

You choose the methodology, the platform builds it

VaultSpeed models and generates for 3NF, Data Vault 2.0, star schemas, ELM and other approaches. The opinions live in the skills you install, so your team builds the way it already works.

3NF

Normalized integration models generated from the context store, with lineage recorded as they are built.
3NF

Data Vault

Hubs, links and satellites from the VaultSpeed Data Vault Generator, for teams that want a vault.
Data Vault

Star schema

Dimensional models for the gold layer and for warehouses built that way, facts and dimensions generated from the same context.
Star schema

ELM and more

ELM and other modeling approaches, delivered as skills with pattern libraries you can review and extend.
ELM and more
The data model committed and pushed to Git from an agent session

Lifecycle and branching

Versioned like code, without becoming code

Every change to the data model is versioned in the graph and reversible. Branches and releases let domains move at their own pace. Governance tools stay the master: VaultSpeed takes in taxonomies, ontologies, ownership and sensitivity from Collibra, Alation, Atlan or Unity Catalog and translates them into what gets implemented.

Branches and releases

Each domain works on its own branch of the graph and releases on its own schedule, on one shared data model.
Branches and releases

Every change tracked and reversible

The data model and its mappings are versioned in the graph, the generated code in Git, so any change can be reviewed and rolled back.
Every change tracked and reversible

Governance tools stay the master

Definitions, ownership and sensitivity come in from Collibra, Alation, Atlan, Unity Catalog or Horizon Catalog, and sync back.
Governance tools stay the master

Where the data model starts

Convert what you have, or draft what you need

Convert from Erwin or PowerDesigner

Import the existing data model into the context store and keep modeling it with an agent.
Convert from Erwin or PowerDesigner

Start from an industry model

FIBO, ACORD, HL7 FHIR, BIAN, ISA-95, SID, HEDM and more, loaded as immutable business definitions to build on.
Start from an industry model

Build the enterprise or canonical model

Draft it with AI from the definitions already agreed on, and keep every domain on the same data model.
Build the enterprise or canonical model

Align to open standards

ODCS, the Open Data Contract Standard, and ODPS, the Open Data Product Standard, for the data product context.
Align to open standards

The data model is where the context lives.

Build it well and every use case, data product and hard question, asked by a person or by an agent, gets a better answer.

campaigninvoiceaddressregionsupplierskupaymentplacescontainscustomer_idemailsegmentorder_dateamountunit_pricebusiness keysource: crmpii flagcustomerorderproduct

The handoff

Engineers query the graph from the harness they already use

The context store is also the handoff between modeling and engineering. The modeler builds the graph, and engineers query it from the coding harness they already use, whether that is Claude Code, Cursor, Snowflake CoCo or Databricks Genie: any agent can query the context store over MCP. Engineers keep their tools and their way of working, and every agent they run starts from the same definitions. A person can ask the store questions in natural language too.

Claude CodeCodexCursor and other IDEsSnowflake CoCoDatabricks GenieVaultSpeed agentsAny agent over MCP

How the code is generatedThe context store in detail

From data model to code

Deterministic where the pattern allows, agentic where it does not

Technical teams need to know what produced each line of code, and VaultSpeed keeps that distinction explicit. Rule based generators or decision models such as JEV produce the same code from the same data model on every run. Where no rule set exists, an LLM writes the code, grounded in the context store and reviewed on every run.

customertablecustomer_idorderemailsegmentsource: crmpii flagbusiness keyorder_dateCONTEXT STOREtables, attributes, keys and lineage in one graphRule based generatorNO LLM IN THE CODE PATHDecision model, JEVSAME INPUT, SAME CODESkill run by an LLMREVIEWED IN GITHUB_CUSTOMER.SQLcreate table dv.hub_customer asselect distinct md5(customer_id) as hk_customer, customer_id, -- business key 'crm' as record_source, load_dtsfrom stg.crm_customer;DDL, SQL or dbt, workflowsone reviewable change in your Git
customertablecustomer_idorderemailsegmentsource: crmpii flagbusiness keyorder_dateCONTEXT STOREtables, attributes, keys and lineageRule based generatorNO LLM IN THE CODE PATHDecision model, JEVSAME INPUT, SAME CODESkill run by an LLMREVIEWED IN GITHUB_CUSTOMER.SQLcreate table dv.hub_customer asselect distinct md5(customer_id) as hk_customer, customer_id, -- business key 'crm' as record_source, load_dtsfrom stg.crm_customer;DDL, SQL or dbt, workflows, in your Git

Rule based generators

The Data Vault Generator turns the data model into DDL, transformation code and workflows through rules and templates. No LLM in the code path, so the same data model gives the same code on every platform.
Rule based generators

Decision models

Decision models such as JEV encode the pattern choices explicitly. Same input, same decision, same code, every run.
Decision models

Other methods run as skills

Pattern libraries are embedded in skills. Install them from the marketplace or build your own skills and plugins. An LLM generates from the skill and the context store, and the output goes through Git review like any other code.
Other methods run as skills
Metadata, not code

Data engineering that works from a shared metadata graph is robust, maintainable and auditable, because the design and business logic live in metadata, not in code.

Customer story · Public sector

VaultSpeed lets us go from data models to impact faster, cleaner, and with more confidence.
Investissement QuébecRegulatory and compliance reporting and dashboards on SnowflakeRead the story

Talk to us

Start with the data model you have

An Erwin export, a PowerDesigner file, or the enterprise model nobody has written down yet. In the first working session we load it into the context store and model the first domain with you.