In my previous role at Planet, I built the data infrastructure to capture metrics for satellite imaging collection for the largest fleet of Earth Observation satellites in the world. At the time I worked there, Planet operated roughly 200 EO satellites taking 3 million images and downlinking 25TB of data every day, creating and selling data products based on that data. In my first role, I was tasked with designing a system that would alert us to any issues with image collection. This was a critical, albeit internal, use case. Imagery was our revenue. If it went down, our product would fail. 

5 years ago, this is how I did it.

Then: Python scripts, SQL, and dbt

My first job was to get data from internal APIs and other telemetry systems. I put everything in the company's existing cloud relational database, which was relatively easy. But then I had to make sense of it. To do that, I needed a system to transform table rows into metrics such as collection latency, coverage, and fulfillment for any image our customers ordered us to take. 

I started this project in the middle of a transformational era in data. The data engineering community was moving away from Spark and Hadoop to the Modern Data Stack. But I had not subscribed to any data influencer Substack yet, so I was totally unaware of what was going on. This project was my data engineering journey.

So, I started scripting data pipelines in Python alongside SQL files, running the executions with cronjobs. I was happy with my code, but the complexity began to grow, and a Frankenstein was born. 

I discovered scheduled queries in my relational cloud database, and while it seemed easy to use and manage, I didn't migrate to it. The lack of version control was a no go for me. Then, dbt came to the rescue. Our team started using dbt and was able to organize and manage data pipelines using software development workflows such as CI/CD and version control. 

At that moment, we started exposing our metrics to internal teams, and we quickly realized the need for a proper Anomaly Detection System. While researching approaches, I stumbled upon an insightful Tinybird blog post that helped me develop some foundational queries. These queries were applied to all our metrics. Using dbt’s macros and OS packages, a lot of copy and paste, and many, many CTEs, we ended up with a decent anomaly detection system. Of course, we then had to think about how to display those metrics through dashboards and an alerting system, but that is a story for another time. 

So, the long story short? I spent many months creating a complex data pipeline, combining my expertise in geospatial data, data engineering, Python, SQL, dbt, and more to create a bit of a Frankenstein internal tool. Did it work? Yes. But it also took a long time to build and perfect, and in the end, it wasn't even something core to our product, our revenue generator.

Now: prompting

Around the time I left Planet and joined Tinybird, I had just started using smart autocompletes and copilots in my IDE. That was a glimpse of what was coming: The era of prompt engineering and AI.

Prompting, (which I guess everyone now calls "vibe coding"), is how you tab your way to a functional product really quickly. Whether you're using a vibe code tool or just prompting in your IDE, the result is the same: rapid prototyping. Of course, vibe code isn't perfect, and many are discovering some… problems with the approach. 

The net of it is that vibe code is great for toys, but not so great for prod. As I look online, I see a lot of people using open source vibe coding tools to build toys: games, clone apps, etc. But recently, I discovered a better and more impactful use for vibe coding: building internal tooling.

With the advent of AI-assisted IDEs and open source vibe coding tools, I decided to reimplement my anomaly detection system. How long might it take to rebuild it using modern tools and AI? The short answer: a day.

Instead of spending weeks configuring multiple tools, I achieved a working prototype in a single day using Tinybird with a few simple prompts. 

Here's my workflow:

 Use tb create with a prompt to get a v0 of my anomaly detection data pipeline:

This command automatically generates Tinybird data files that define the data sources and endpoints that would become my anomaly detection system. Here, for example, is the raw data source Tinybird created to hold all the satellite image collection information.

It also created a materialized view that aggregates satellite image data by hour.

It created a materialization pipe to populate the data source:

And an endpoint to detect anomalies in the satellite image collection latency using a z-score algorithm on a 10-hour rolling window.

That was a 1-shot. All this, with just a prompt.

Of course, I wanted to test it. So, I ran tb mock with a prompt to generate fake data as close as possible to real data:

This created an .ndjson file with 10 rows, like this:

One cool thing about tb mock is that you can see what it has done behind the scenes. It creates an SQL file, in this case satellite_images.sql, which shows you how the mock data gets generated:

Magic, right? 

At this point the structure of the data project resembled something like this:

All that was left to do was deploy it to the cloud with tb deploy --cloud, and the anomaly detection API was live.

Four commands, two prompts, and one functional anomaly detection API that, in my past life, took me weeks, if not months, to build. This is the way. If you're exploring ai vibe coding free options, Tinybird's free tier is a great place to start.

All that's left is to use your favorite vibe code tool to build a Next.js app or some other app to visualize and alert. With ai vibe coding free tools becoming more accessible, I won't show you how to do that. There are a million examples on Twitter.