Last week, Philly, Jason, and I hopped on a livestream to show off what we've been building. It's been about six months since we really doubled down on AI agents, and we finally got to reveal the notebook interface we've been working on.

I've got to say, it felt good to finally show this to people.

How We Got Here

We started the stream talking about how the three of us even met, which is a story I love telling. Philly reached out to me on LinkedIn years ago to basically pick a fight about automated machine learning. I was at DataRobot at the time, working on COVID forecasting for the US government, and we were running these Jupyter notebooks that took an hour to execute from start to finish.

The thing about notebooks is they're great for iterating quickly, but they have real problems. Collaboration is a nightmare… Dependencies break, version control is a mess, all of that. And Philly and Jason were dealing with the same stuff from rural Ireland, wondering how these big companies with all their resources were still struggling with basic tooling.

That confusion led to a meeting at a pub on Southampton Street in Dublin. We had a few pints and started sketching out what would eventually become Zerve. Jason even told the story about explaining to our head of engineering how notebooks can make 1 + 1 equal 7, and the guy literally didn't believe it could work that way.

What We Actually Built

So here's what we showed on the livestream this week.

The Notebook Interface

You get the traditional notebook view everyone's used to. Code cells, inline outputs, the whole thing. But it runs in the cloud, so you get the same results every time. And multiple people can work together at the same time, (think Google Docs for data).

We added some nice touches:

• AI-generated descriptions for each code block

• A gallery view so you can see all your charts in one place

• You can toggle between the notebook view and our DAG canvas view

• Full-screen modes when you need to focus

The notebook and the canvas are connected. Click a block in the canvas and you jump right to that cell. You can rewire your code dependencies just by dragging and dropping in the DAG. And when you run everything, blocks that don't depend on each other execute in parallel.

The Agent

This is where it gets fun. Our agent is built specifically for data work. It understands data workflows in a way that general coding assistants don't.

Jason did a demo during the stream. He uploaded the Titanic dataset (yeah, that one) and just asked the agent to do some exploratory analysis with Google charts. The agent made a plan, built multiple blocks, and churned out violin charts and survival analyses. Jason didn't touch his mouse.

The key is context. The agent sees your data, your code, your outputs, your database connections. You don't have to keep explaining what you're working on.

You can work two ways with it:

• Give it broad prompts for comprehensive analysis

• Use the block-level Ask AI for quick tweaks

And you can jump in anytime you want. The agent runs on its own, but if you want to work on something else, just click on a block and start coding. It won't interrupt you.

Getting to Production

This part is straightforward. Set your notebooks to run on a schedule and they'll retry if they fail. Deploy APIs. Turn notebooks into apps.

During the demo, we showed how you can just ask the agent to turn your notebook into an app. Half a minute later, done. You can set up environments with the packages you need and reuse them across projects. Everything gets its own URL and scales automatically.

How This Changes Data Work

I've been thinking a lot about how this changes the way we work. The friction is basically gone. When trying something new takes seconds instead of an hour, you just try more things.

I told a story on the livestream about my son asking for help with a 3D printing problem. He needed some edges rounded in a specific way. I took his file, uploaded it to Zerve, and asked the agent to handle it. It worked.

That's the thing - If you know a concept exists, even if you've never implemented it, you can just describe it and the agent handles the code. Need SHAP values? Want to check for multicollinearity? Ask Zerve.

All three of us are doing way more analysis now than we were a year ago. Jason said it best: "You don't have to make such a commitment to be able to see if an idea will work or not."

Questions People Asked

We got some good questions during the stream.

Someone asked how this is different from Jupyter or Google Colab. I explained that Jupyter was built as a classroom tool and runs locally in memory. Colab is just Jupyter online and somehow makes collaboration even harder. We built Zerve for actual work.

Another question was about security. You can self-host Zerve in about six minutes. Everything runs in your infrastructure. We built it this way from the start because we knew organizations like NASA would need it.

Where We Go From Here

Near the end, I asked Philly and Jason about the next five years. We think AI should stop waiting for instructions. Eventually systems will just fix problems and tell you about it after. We're already seeing this change how people work.

One viewer predicted that in five years, demos like ours would be done by AI, and the audience would be AI representatives creating summaries. We thought he might be right. Not sure yet how I feel about that.

Try It Out

If you want to see the whole thing, watch the livestream. We go into more detail about the origin story and show more of the product in action.

You can sign up for free and start building. We give you plenty of credits to get going.

And yeah, the story about the Dublin pub is worth hearing in full.