Chai With JaiGPU Study Hall
Menu
← Read the essayCHAI WITH JAI · LABS / 01

Snowflake · Unstructured data · Evaluation

Unstructured Data
Modeling on Snowflake

A guided Day 1 field lab: build two PDF-extraction paths, test both against independent SEC facts, and make the architecture decision a senior analytics engineer is expected to defend.

0 / 18 verified

01 · Frame the decision / Trust test

Understand

Would you ship this number?

Your move: Notice the difference between a value that looks plausible and one that is ready for a data product.

EXPECTED RESULT

Before opening Snowflake, you can name at least one obligation that the number alone cannot prove.

Do this in order

  1. Read the model output and make a verdict before revealing the answer.
  2. Explain what evidence would change your verdict.

The annual report shows “Net sales 383,285.” The model returns 383,285 USD. What is the first failure?

In one sentence, why is 383,285 USD not yet trustworthy?

Check your diagnostic answer first. Record evidence before verifying.
WHY THIS STEP EXISTS

The long tail begins with ordinary document details—scale notes, periods, table layout, and evidence—not exotic model behavior.

IF THE RESULT DIFFERS

If you chose the model output because it looked reasonable, inspect the statement header. Ask what unit and multiplier govern the entire table.

Evidence boundary

This companion follows the inspected DataExpert course repository at commit c03992d and the Day 1 lecture and lab recordings. The bundled starter is the baseline because it preserves the evaluation skill without external-download variation.

Local notes and progress stay in this browser. Never place passwords, API keys, access tokens, or credentials in worksheets, screenshots, notes, or the decision memo.

Return to the cover ↑