Data architect · Louisville, KY · remote-friendly

Ask your data a question. Get a governed answer.

A talk-to-data assistant I built at work, with access controls built in.

I design data architecture for analytics and AI, and I help teams adopt AI in their everyday data work.

Synthetic data · rule-based demo, no LLM · try it live
Resume See the demo
10
early users
proof of concept, launched May 2026, built at work
115/124
evaluation questions correct (~93%)
synthetic-data demo; the other 9 were refusals or clarifications, none wrong
16
automated checks in a worked example
AI-assisted data engineering, synthetic data; caught 6 of 6 seeded defect types
Scroll

Overview

The whole story in 40 seconds.

The looping clip is silent. The full version runs 40 seconds, has music, and every line is captioned on screen. Demos use synthetic data.

Read the text version

Ask your data a question. Get a governed answer. A talk-to-data assistant I built at work, with access controls built in. Proof of concept: 10 early users; 115 of 124 evaluation questions correct on synthetic data; 16 automated checks in a worked example. In the demo, a question in plain English is answered with the query shown, and a growth-by-category question is computed from governed metric definitions. A Sales analyst asks for margin and is refused and logged, which is the correct answer. I work with AI across the whole data engineering workflow, and I'm helping others on my team adopt it. In the worked example on synthetic data, defects were seeded on purpose: 6 of 6 defect types were caught by 16 checks. The assistant's draft v1 failed 5 of 9 model checks; after the rewrite 7 of 9 pass, and one stays red on purpose because it flags a defect in the source data.

Overview, 40 seconds, with sound

Music is original and synthesized for this video. Captions are on screen. Press Esc to close.

The problem

Everyone has data questions. A trusted answer is the hard part.

“What were total net sales in 2025?” “Top 3 brands by net sales in 2025” “Net sales growth by category, 2025 vs 2024”
One definition per metricneeded
Access that follows the roleneeded
Answers you can checkneeded

The design

Ask in plain English. Answers come from governed data.

01People business users,by role 02Plain-English chat questions in,answers out 03Guardrails access control,refusals, audit log 04Curated data sets governed metricdefinitions 05Answers with the queryshown, and checkable GOVERNED: ACCESS BY ROLE, ONE DEFINITION PER METRIC 01Peoplebusiness users, by role 02Plain-English chatquestions in, answers out 03Guardrailsaccess control, refusals, audit log 04Curated data setsgoverned metric definitions 05Answerswith the query shown, and checkable

Guardrails and evaluation

Access is limited by role. Failures are measured, not hidden.

Guardrails

A refusal, from the demo

What is gross margin by brand in 2025?
Refused
Your role (Sales analyst) is not permitted to query Gross margin %. Ask an owner of that data for access.
Role comes from the session, never from the question.
Governance and safety →

Evaluation

124 questions, gold answers

  • Gold answers come from hand-written SQL, run independently of the assistant.
  • Failures are labelled, written up, and published on the evaluation page.
  • Every request, answered or refused, goes in an audit log.
Evaluation and failure analysis →

Results

A small proof of concept, measured honestly.

10 early usersproof of conceptlaunched May 2026built at work

The evaluation numbers come from the synthetic-data demo, not from the work deployment. Some of the questions were used while tuning it. On the initial version, the untuned stress set scored 15 of 30.

115of 124

correctrefused or asked to clarify (0 wrong answers)

Go deeper

Working demos, on synthetic data.

  • I work with AI across the whole data engineering workflow, and I'm helping others on my team adopt it.
  • Built a natural-language analytics assistant over curated data, with access controls, as a proof of concept.
  • Leading the architecture for an enterprise financial-planning data model.
  • The projects below are self-directed and use synthetic data.
Talk-to-data demoAsk questions as different roles. See the query behind each answer.Open → Governance and safetyRole policy, audit log, and a red-team suite that runs in your browser.Open → AI-assisted data engineeringA worked example on synthetic data: seven steps with a human decision at each, computed checks, and a small evaluation.Open → ResumeExperience, skills, and contact details.Open → ArchiveEarlier projects, kept online.Open →

Roles I'm looking for: AI enablement, data architecture, and analytics leadership.

Watch the 40-second overviewEmailLinkedIn© 2026 Maxwell Mattox