DEV Community

Cover image for We Built 100 AI Analytics Tasks. Then We Asked: Who Verifies the AI?
šŸŽ“ Navid Ishaq
šŸŽ“ Navid Ishaq

Posted on

We Built 100 AI Analytics Tasks. Then We Asked: Who Verifies the AI?

Three repositories. One analytics platform. A hundred AI tasks. And one question that changed the project.

At DeTLeng, we built a working e-commerce analytics case study using the public Brazilian Olist dataset. The architecture was solid: BigQuery for processed analytics, FastAPI for approved SQL tools, OpenAI for natural-language explanations, and GitHub Pages for the website.

We even consolidated three separate repositories—the website, AI backend, and BigQuery analytics tooling—into one maintainable codebase.

Then all 100 AI Task cards started answering business questions. Revenue? Yes. Customer trends? Yes. Delivery performance? Yes.

Everything looked ready for a portfolio demo.

But I kept coming back to one question:

An AI answer sounds convincing. Can the visitor verify the numbers without trusting the AI?

From ā€œAsk AIā€ to ā€œVerify with SQLā€

That question became DeTLeng ProofSQL.

Here's the experience we built:

  1. Select an AI Task and request its live BigQuery-backed analysis.
  2. Read the original AI explanation.
  3. Click āš– Verify with SQL under the answer.
  4. Watch DuckDB-Wasm execute an approved SQL query inside your browser against a versioned analytics snapshot.
  5. Inspect the comparison: MATCH, MISMATCH, or NOT VERIFIABLE, with data checksums and SQL provenance.

The verification click does not trigger another BigQuery query or an OpenAI request. It's a separate calculation path, using SHA-256-checked CSV snapshots—not a second AI opinion.

We started with five core KPIs. The first live revenue check reproduced 15,843,553.24 on both sides.

Then we expanded the architecture to 100 SQL mappings and a nine-table snapshot.

The Bug That Made the Project Better

Here comes the part I would absolutely include in an engineering interview.

Before presenting the project, I tested tasks beyond the initial pilot. Their AI answers worked—but many Verify buttons were missing.

The backend was deployed. The SQL definitions existed. The data was available.

The actual problem? A frontend condition still expected the pilot's verification.metrics field, while the expanded backend also returned verification.records.

A small, targeted JavaScript compatibility fix restored the missing path. No dramatic rollback. No rebuilding the platform. Just investigation, regression tests, a reviewed Git diff, and another live test.

Recent browser checks then showed Task #92 matching across 24 rows and 120 numeric comparisons, with Tasks #96 and #100 also returning matching numerical evidence.

Does that mean every one of the 100 tasks has been individually certified? No. It means the 100-task verification architecture is in place and we have documented live evidence across multiple types of analysis. That distinction matters.

Don't Just Read the Case Study. Test It.

If you're a developer, analyst, recruiter, or simply curious about practical AI engineering, I'd love you to try the system yourself:

Pick a task. Ask the AI. Open ProofSQL. Look at the numbers and the SQL. Tell me what works—and what should be better.

This was human-led, AI-assisted engineering: AI helped with architecture, code, tests, and debugging; I kept challenging the outputs, running the deployments, and checking what actually worked in production.

The lesson: A working AI demo is a beginning. Reproducible evidence—and the willingness to find your own bugs—make it an engineering project.

And when the SQL matches? As we say: LUDDI HAI JAMALO! 🄁

About the Author

Muhammad Naveed Ishaque is a multidisciplinary thinker, educator, and founder of the DeTLeng Ecosystem, exploring AI, data, technology, business, and human-centered innovation through continuous learning and practical research.

🌐 Explore DeTLeng: www.detleng.com

Top comments (0)