Data Platform & Vibe Coding: Passing Trend or Real Revolution?
You’ve probably seen the term “vibe coding” pop up over the past few months.
The concept, popularized by Andrej Karpathy, describes a way of building software that is quite different from what we were doing until now: you let AI generate a large part of the code, you iterate quickly, and you prioritize the result over the perfect implementation.
In short:
“Does it work? OK, let’s move on.”
In the world of data, this approach is also starting to appear. You assemble managed services, generate scripts with an LLM, connect a few tools… and a data platform can emerge very quickly.
Some are already calling this “Vibe Data Engineering”.
But is it a good idea?
Is it the future… or the next massive technical debt?
Let’s look at both sides.
What is “Vibe Data Engineering”?
Imagine the situation.
You need to build a pipeline that ingests logs, cleans them and exposes them in a dashboard.
Classic approach:
architecture → design → review → implementation.
“Vibe” approach:
You paste a log sample into Claude or ChatGPT and write:
“Write me a Python script that parses this and sends it to BigQuery.”
A few seconds later: you have a script.
You test.
It works 80%.
You ask it to add error handling.
You tweak it a little.
Meanwhile:
you deploy Airbyte
you connect the source
you expose the data in Metabase
One morning later, you have a working pipeline.
Not necessarily perfect.
But usable.
And that is exactly what the “vibe” is: moving fast, hacking intelligently, assembling building blocks.
Why it is so appealing
1. Speed
Business needs change quickly.
With AI, a new KPI can go from idea to dashboard in just a few hours.
In some contexts, that is a huge advantage.
2. Less technical friction
No more spending 20 minutes on Stack Overflow trying to remember the right syntax.
AI generates a codebase that can be used immediately.
You focus more on the data and the use case.
3. Testing ideas quickly
Want to try a new data source?
Instead of launching a structured project, you build a quick little pipeline.
If it is useless, you delete it.
The cost of experimentation becomes very low.
4. More autonomous teams
With AI assistants, some analysts or PMs can create simple pipelines themselves.
It is not perfect… but it makes things much smoother.
5. Well suited to early-stage phases
When a project starts, the requirements are still unclear.
In those moments, moving fast is often more useful than designing a perfect architecture.
The risks — and they are real
1. Invisible technical debt
Generated code often works…
but it is not always clean or maintainable.
Six months later, you discover:
incomprehensible scripts
few or no tests
no documentation
And nobody wants to touch the pipeline anymore.
2. The illusion of productivity
You feel like you are moving very fast.
But sometimes you then spend a lot of time fixing subtle bugs introduced by AI.
And problems often appear at scale.
3. Governance and security
An LLM does not know your company’s internal constraints.
It does not know that:
this column is sensitive
this dataset is subject to GDPR
this export must never leave the network
Without guardrails, the risk exists.
4. Loss of understanding
If everything is generated by AI, you sometimes understand less about what you deploy.
And when a pipeline breaks at 3 a.m…
you need to know how to debug without AI.
5. Scaling up
Vibe works well:
alone
or in a small team
But when a platform serves dozens of teams, the lack of standards quickly becomes a problem.
6. Cloud costs
An AI-generated SQL query can easily scan much more data than necessary.
And in systems like BigQuery or Snowflake…
the bill can climb quickly.
So… what should we do?
As is often the case, the answer is not binary.
Vibe is not necessarily bad.
But it depends on the context.
A few useful questions:
1️⃣ How mature is the team?
Experienced engineers can use AI as an accelerator.
A very junior team mostly risks piling up problems.
2️⃣ How critical is the data?
Internal exploration ≠ financial reporting.
3️⃣ What stage of the project are you at?
0 → 1: speed
1 → 100: structure
4️⃣ What guardrails exist?
For example:
code review
pipeline templates
automated tests
data lineage
cost monitoring
Vibe works much better within a clear framework.
To conclude…
Vibe Data Engineering is neither heresy nor a miracle solution.
It is simply a new way of working, made possible by LLMs.
Used intelligently, it is an incredible accelerator for prototyping and exploring.
But to build a sustainable platform, structure remains essential.
Deep down, the right approach is probably:
Structure + Vibe.
Solid foundations…
and the freedom to experiment within them.
Are you more:
team vibe
team structure
or a mix of both?