inflearn logo

Vibe Modeling - How to make AI agents perform data analysis and modeling

If you ask an AI to "build a model with this data," the first few lines might look plausible, but as soon as things get slightly complex, it starts using the wrong columns, introduces data leakage, and results in a model that performs worse despite a higher score. This course does not teach you better prompts. Instead, it covers how to directly design a loop (harness) where an agent understands data, plans, writes code, validates its own results, and backtracks when wrong. After manually walking through the five-piece loop of DS-STAR, you will assemble components one by one, from data profile cards, hooks, and skill hierarchies to independent validators and ablation.

1 learners are taking this course

Level Beginner

Course period Unlimited

Python
Python
Machine Learning(ML)
Machine Learning(ML)
Pandas
Pandas
mlops
mlops
LLM
LLM
Python
Python
Machine Learning(ML)
Machine Learning(ML)
Pandas
Pandas
mlops
mlops
LLM
LLM

What you will gain after the course

  • How to design a loop (harness) that ensures AI agents complete data analysis to the end

  • How to write data profile cards that prevent agents from misunderstanding data

  • How to turn rules into enforcement mechanisms rather than just documents using hooks and skill layers

  • You cannot verify your own code - Designing independent verifiers and judges

  • An iterative routine that actually boosts performance through ablation, ensemble, and leakage testing.

"Build a model with this data" — Why it doesn't end with just that one line

If you've ever assigned an analysis to an AI, you'll know. The first few lines are surprisingly plausible. But as soon as things get a little complex —

· It sneakily uses columns that shouldn't be used

· Future information gets mixed into the training data (leakage)

· The metrics went up, but the actual model got worse

· They confidently present a wrong answer, saying, "It's all done."

This is not because the model is stupid. It's because you asked it to do everything at once.

■ This is not a prompt engineering course

I won't be teaching you better prompts. Instead, we will design a loop that allows the agent to run autonomously—a 'harness.' A five-part loop that understands data first → makes a plan → writes code → validates its own results → and goes back if it's wrong. After manually running through the very structure used by DS-STAR, you will assemble it to fit your own tasks.

■ You will build these yourself

· Data Profile Card — The first step to ensuring agents do not misunderstand the data

· Hook — A mechanism that enforces rules instead of just documenting them.

· Skill Hierarchy — Including everything is the same as including nothing. Make it so only what's necessary is activated.

· Experiment Ledger — A recording structure to avoid making the same mistake twice

· Independent Verifiers and Judges — You cannot verify your own code.

· Routers and failure exits — how to stop without getting stuck in an infinite loop

· Ablation · Ensemble · Leakage Checker — The final stretch to actually boost performance

■ Why now

Models are already smart enough. These days, models even ask follow-up questions. The problem is that there is no one to answer them. A gap is widening between those who know what to ask and what to verify with an agent, and those who do not.

Upon completing 31 lessons, you will walk away not with a slide deck, but with a loop in your hands.

※ The videos for this lecture were produced using AI voice (TTS).

Recommended for
these people

Who is this course right for?

  • A data analyst who asked AI for analysis but only received plausible-sounding wrong answers

  • A junior ML engineer whose performance no longer improves after the baseline.

  • Developers who are familiar with "vibe coding" but couldn't apply it to data and modeling tasks

  • MLOps and platform managers who need to design agent pipelines themselves

Need to know before starting?

  • Experience in handling data using Python syntax and pandas

  • Experience using AI coding tools like Claude Code or Cursor at least once

Hello
This is sof

118

Learners

12

Reviews

1

Answers

5.0

Rating

2

Courses

SoftData is a team that creates data-related educational content.

Curriculum

All

31 lectures ∙ (1hr 57min)

Course Materials:

Lecture resources
Published: 
Last updated: 

Reviews

Not enough reviews.
Please write a valuable review that helps everyone!

Similar courses

Explore other courses in the same field!

Limited time deal

$13,860.00

30%

$15.40