Skip to content
Mohd Khizer Hayat

← All work

CS-02 · 2026 · ongoing

Self-serve analytics with a governed LLM data layer

Letting people ask questions in their own AI tools and get trusted answers, by constraining the model to vetted warehouse definitions instead of improvised SQL.

In progress. This is emerging work, presented as it stands today.

Where the governance sits
  1. A questionAsked in the person's own AI tool
  2. Model-context layerExposes a curated slice of the warehouse
  3. Governed viewsVetted definitions
  4. WarehouseThe underlying data

The model never touches raw tables. Every answer is built from definitions that were agreed and vetted before the model arrived.

The problem

Every number that isn’t already on a dashboard means a business user waits in a queue for an analyst to build a custom report. That doesn’t scale, and it puts the analytics team on the critical path for routine questions. The goal: let users ask questions in their own AI tools and get trusted, governed answers without an analyst building something each time.

What I am building

A governed interface that lets LLM-based tools query the data warehouse through vetted definitions rather than improvised SQL.

  • A model-context layer (MCP-style) that exposes a curated, governed slice of the warehouse to AI tools.
  • A lightweight delivery interface built on a rapid app framework, with an explicit principle of not carrying legacy reporting pain points into the new surface.
  • A weekly working cadence to converge on scope and on the governance rules that keep answers trustworthy.

My role

Contributor on a small cross-functional team driving the initiative. This is collaborative and still in progress. My part is the warehouse and governance side: making sure the layer the model queries returns trusted, consistent results.

Architecture and trade-offs

  • Flexibility vs trust. An LLM that can write arbitrary SQL against raw tables will eventually return a confidently wrong number. The design constrains the model to a governed view and definition layer, so answers come from vetted logic.
  • Reuse over reinvention. The AI layer sits on an existing governed semantic layer, so it inherits definitions that are already agreed on rather than inventing new ones.

Impact

Target outcome: shorter turnaround on ad-hoc requests and fewer analyst hours spent on one-off pulls. Metric pending: there is no post-launch measurement because the work is pre-launch.

The hardest part

Governance, not model wiring. Making sure an LLM answers from trusted, consistent definitions rather than improvising is the core challenge, and it is why the governed view layer underneath matters more than the model on top.

Skills shown

LLM and AI tooling, MCP-style integration, data governance, semantic-layer design, rapid prototyping, cross-functional collaboration.