GW
HomeWorkAboutNotesStackLab
Experience atOmnicom Group

Driving financial systems and automation across global teams.

Builder ofEnterprise AI Systems

AI agents, audit workflows, and decision intelligence at scale.

Based inNew York City

Building systems for enterprises around the world.

HomeNotesWorkAboutStackLabPartners
XLinkedInGitHubEmailllms.txt
© 2026 · New York City
All enterprise work

Platform & data case study

Governed analytics and natural-language access.

A permissioned analysis layer that lets finance users ask questions over approved views and inspect supporting context.

Databricks logoDatabricksUnity CatalogAI/BI
Ask the project copilotGrounded in this public-safe case study.

Business problem

What people needed to solve.

Finance teams want faster answers, but an unrestricted natural-language route into sensitive data can return unverifiable results or bypass access controls. The layer has to enforce what each user can ask and see, tie every answer to an approved view, and let the user inspect and challenge the result.

Decision frame

Questions the work needed to answer.

  • What can this user ask and see?
  • Which approved view supports the answer?
  • How can a user inspect and challenge the result?
Interactive operating flowSelect any node for context or ask the AI guide.

Selected Operating layer

Governed analysis layer

Connected approved planning data to a governed analytics layer where user access boundaries are applied before the natural-language interface, and answer paths keep the source and result context available for review. Output is treated as decision support, not an unverified system of record.

Implementation

From the business problem to a working operating model.

Connected approved planning data to a governed analytics layer where user access boundaries are applied before the natural-language interface, and answer paths keep the source and result context available for review. Output is treated as decision support, not an unverified system of record.

  1. 01

    Selected approved data domains and consumption views.

  2. 02

    Applied user access boundaries before the natural-language layer.

  3. 03

    Designed answer paths that keep source and result context available for review.

Controls, approvals, and delivery constraints

The work around the work.

Enterprise systems change only when data, access, testing, ownership, and evidence move together.

  • Natural-language output is decision support, not an unverified system of record.
  • User permissions and source context stay visible.
  • No prompt, model, or internal-data implementation detail is disclosed publicly.

Prior art

How leading teams approach this.

Public references for the same class of problem, so this work can be read against how other organizations are building it.

Databricks

AI/BI Genie is now Generally Available

Databricks Genie grounds natural-language answers in Unity Catalog as the single source of truth for data and semantics, enforces existing access policies, and supports verified answers plus per-question activity monitoring, the governance model this layer follows.

Anthropic

Claude for Financial Services

Anthropic routes Claude to finance data through governed connectors with full audit trails so analysts can trace an answer to its source, the inspect-and-challenge requirement built into this layer.

Morningstar

Morningstar Credit Analytics Launches AI Access to CRE Surveillance and CMBS Analytics

Morningstar's natural-language access delivers data only within each user's existing entitlements, the access-boundary-before-the-NL-layer design used in this case study.

Public-safe outcome

Illustrative of plain-language analysis with clear data-access boundaries, visible source context, and no fabricated answers, with implementation detail such as prompts and models kept out of the public example.