The email was about three sentences long. Jakub Jurovych introduced himself, pointed to work he had already made, included a GitHub link and a screenshot, and sent the note to Mozilla. There was no grand career pitch. There was a small piece of evidence attached to a clear interest in developer tools. Roughly six months later, he was flying to San Francisco to turn those ideas into Firefox DevTools. In Jurovych's telling, one email changed the direction of his life. The more durable lesson sits inside the attachment: when he wanted a door to open, he brought something that worked.
That instinct now runs through Deepnote, the company he founded in 2019. Its first persuasive demo was almost as spare as the Mozilla email. Jurovych opened two browser windows on one laptop, typed a few characters into the left side, and watched them appear in the right. Two cursors could inhabit the same data notebook. An awkward, mostly single-player document had become a shared surface. Investors could understand the product before anyone finished explaining the market.
Seven years, two announced funding rounds, one acquisition, and an open-source release later, the demo's question has widened. What happens when the other cursor belongs to an AI agent? Jurovych starts his answer inside the notebook, where the work already lives. The notebook can hold the question, code, data connections, explanation, permissions, history, and the judgment required to decide what happens next.
Build for the people who build
Jurovych grew up in Zvolen, Slovakia, and started making websites as a child. By around 17, he said, programming was already paying him. University initially looked optional. He enrolled in informatics at Masaryk University in Brno on a one-semester trial with himself. The experiment lasted three years, ended with a bachelor's degree in Applied Informatics, and placed him among the top one percent of the program by grade average.
One former teacher remembered a student who did not seek attention but arrived organized, focused, and unusually prepared. The description fits a career built less on public theater than on carefully framed technical artifacts. Jurovych applied to Oxford and Cambridge, was accepted to both, picked Cambridge, then deferred for a year to finish work. At Cambridge he studied advanced computer science, concentrating on machine learning and human-computer interaction. He also learned to play polo, a stray detail in a biography otherwise crowded with code, startups, and plane tickets.
The connecting tissue was human-computer interaction: the study of how a tool changes what its user can think and do. Jurovych calls the impulse “building for builders.” At Mozilla, the builder was a web developer inspecting a page. At Cambridge, it was a researcher exploring a problem. Later, at Deepnote, it would be a data scientist trying to combine code, evidence, and explanation without losing the thread.
“A data science project is not just the code itself - it's the combination of code, data, and execution environment.”Jakub Jurovych, writing at Deepnote
The Hollywood detour that found the real problem
After Cambridge, Jurovych moved to Los Angeles and became CTO of Operam. The company applied machine learning and attribution models to movie marketing, helping studios decide how to allocate advertising budgets. Its work touched more than a hundred films, including campaigns connected to Moonlight and The Shape of Water. The job placed a technically trained founder inside an industry where a model's output had to travel beyond the person who wrote it.
That exposed a familiar failure. Data teams could do sophisticated analysis, but the software around them made collaboration difficult. A notebook might contain the experiment, yet sharing it meant passing files, rebuilding environments, or explaining results in another tool. Add more analysts and the problem expanded. Code was only one piece. Data, infrastructure, narrative, and review all had to move with it.
Operam was growing when Jurovych stepped away from day-to-day work. He retained a stake and served as an adviser, but advertising was not the subject he wanted to spend the next stretch of his life on. The tooling problem was. It had followed him from JavaScript development to academic research to a commercial analytics team. Deepnote began from that repetition.
The company entered Y Combinator's Summer 2019 batch. In early 2020 it emerged from stealth with a $3.8 million seed round from Y Combinator, Index Ventures, and Accel. A $20 million Series A followed in 2022. The money matters because it bought time to pursue a medium, not just a feature: a cloud notebook where teams could connect data, write Python or SQL, edit together, preserve an environment, and share the result through a link.
Two cursors become a company
The first version aimed at data scientists. Usage at universities offered an unexpected distribution path. Deepnote appeared in courses at institutions including Cambridge, MIT, and Harvard. Students could open a browser and enter an already configured project; instructors could see and share the work without turning setup into the first lesson. What looked like an academic beachhead also served the larger thesis: fewer people should be locked out of data work by the ceremony around the tools.
Deepnote gradually widened its vocabulary. A notebook became a project. A project could include several notebooks, shared integrations, environment settings, dashboards, scheduled jobs, and API endpoints. In 2024, the company acquired Hyperquery, a close competitor focused on making analytics legible to decision-makers. Jurovych said the products were philosophically near enough that customers could migrate with little friction. The deal also brought Hyperquery's customers and its Win With Data community into Deepnote.
Mostly individual
Apps + schedules
Review + memory
There is a useful product lesson in that expansion. Deepnote did not discard the notebook as the market changed. It kept asking what the notebook was the smallest unit of. At first, it was an analysis. Then it was a collaborative project. Now Jurovych describes it as executable markdown for an agent: narrative, code, permissions, tools, and environment living in one inspectable place.
The 200-pixel test
When ChatGPT arrived, the easy product response was visible across software: squeeze the existing interface, add a chat panel, and call the result AI-native. Jurovych was skeptical that removing 200 pixels from a product and filling them with conversation would create a tenfold productivity jump. Deepnote moved deliberately while the team looked for interactions that belonged inside analytical work.
The distinction is concrete. A chatbot can propose SQL. An integrated agent can write the SQL into a notebook, run it against an allowed connection, inspect the output, leave a reviewable change, and preserve the path from question to result. The user does not only receive an answer. The team receives an artifact that another person can rerun, challenge, or extend.
Deepnote's 2025 open-source release pushed the same argument down into the format. Traditional Jupyter files were designed before cloud collaboration and AI. Deepnote introduced a human-readable YAML format intended to produce cleaner diffs, represent a multi-notebook project, carry structured metadata, and give agents stable blocks to target. Open formats also answer a question that becomes sharper when software can act on a team's behalf: can people inspect the state, export it, and leave?
“We're building a truly universal computational medium for collaboration between humans and AI agents.”Jakub Jurovych, 2025
A place for the next cursor
In June 2026, Deepnote announced a native connection for a Codex data analytics plugin through MCP. Codex can work within a connected user's existing permissions to search projects, read or write notebooks, run them, and inspect integrations. In August, Jurovych introduced Deepnote Agent Workspace, organizing the product around three building blocks: skills that carry reusable knowledge and access rules, agents that operate in notebooks, and apps that deliver the work to people who may never open one.
The ambition is partly technical and partly organizational. Companies want agents to understand business definitions, trusted tables, recurring procedures, and the exceptions people remember but schemas do not. They can build a separate context layer, or they can treat existing analytical work as the layer. Jurovych argues that the data team is the natural owner because it already turns messy questions into governed definitions and evaluates whether the results hold up.
This makes the notebook sound less like a programmer's document and more like a company's working memory. The notebook remembers how “active customer” was defined, which credentials may touch which warehouse, what the last analysis concluded, and what an agent changed. It can become an app for a colleague, a schedule for a recurring task, or an endpoint for another system. The interface remains open enough for exploration but structured enough for accountability.
Jurovych's career has followed increasingly consequential feedback loops. A browser tool helps a developer see what code did. A film model helps a studio see what advertising did. A collaborative notebook helps a team see what the data says. An agent workspace must help everyone see what software did on their behalf. Each loop gets larger, but the design requirement stays intimate: make the action visible, keep the context nearby, and let a human steer.
The three-sentence email still feels like the right opening image because it contains the whole method. A claim travels farther when it includes a working object. Deepnote's next object is no longer one screenshot or two cursors. It is a shared place where the question, calculation, collaborator, and record can live together. Jurovych is betting that when the next cursor arrives, the notebook will be ready for it.