Explosion

If you've been following us, you probably know that we've always wanted to bring our annotation tool Prodigy into a collaborative cloud environment.* Annotation should be part of the development workflow and labelled data unlocks SO much potential for custom AI solutions – but it's still very tedious. It doesn't have to be, though! While our tools are powerful, a big problem has always been *how* to best use that power for different real-world use cases. This can't really be solved by just a UI. As it turned out, the missing piece was coding assistants and agents: they can guide you through the process, apply best practices, set up annotation tasks and even create data for you. If you're already using Prodigy for annotation, we've put together a page on how our new tool and platform Ellf can help you scale up and automate your data development projects (link in comments). We're still open for beta partners, so if you need help with a project or want to upgrade Prodigy with user management, annotation agents and coding assistant support for custom recipes and interfaces, sign up for the waitlist or get in touch! * If you've been following us even more closely, you might have also seen that we almost got there, but unfortunately ran out of time and money. But we were finally able to bring everything we had originally planned for Prodigy Teams (and more!) to Ellf.

I have known Gor Hayrapetyan since I first became involved in data science in Armenia. He can be considered one of the pioneers of data engineering in the country and has become one of the most influential voices in the local community. I have never met Ines Montani in person, but I have followed her talks at conferences around the world. Over the years, she has become not only one of the most recognizable figures in Python and NLP, but also one of the people most strongly associated with the global PyData community. PyData PyCon Yerevan 2026, July 24–25 Explore the full schedule: https://lnkd.in/gCPzuPCS Get your tickets: https://lnkd.in/ghy4TD3m

Just published a new blog post ✨ Our recent work on NLP workflows got me thinking about Atomic Design by Brad Frost, a methodology for building web interfaces, and how neatly it maps to how we've been approaching applied NLP: atoms (components) → molecules (pipelines) → organisms (applications) → pages (products). The results are systems that are more accurate, cheaper, more transparent, and actually maintainable – instead of one giant LLM call you can't test or reason about. This is exactly the kind of stuff we want to make easier for people! (link in comments)

𝑽𝒊𝒃𝒆 𝑵𝑳𝑷 𝒇𝒐𝒓 𝑨𝒑𝒑𝒍𝒊𝒆𝒅 𝑵𝑳𝑷, A Talk By Ines Montani & Matthew Honnibal, Founders of Explosion 👉 THE FULL TALK IS NOW ON YOUTUBE: https://lnkd.in/g95KUBsg 🧩 Applied NLP’s biggest challenge is translating complex business problems into machine-learning components, a task that traditionally requires deep domain expertise. This talk explores how ideas from AI coding assistants can be applied to NLP, using assistants and MCP servers to help developers build practical, modular, open-source solutions that are small, fast, and fully data-private. The key insight is to use LLMs to build systems rather than be the systems themselves. Just as coding assistants have transformed software development, they can accelerate AI development, reduce expertise barriers, and help teams deliver successful NLP projects.

The video of my talk at PyData London is now live! I discuss the evolution of coding agents, why we should use LLMs to build systems, what's currently holding us back, and how we built a virtual assistant to help developers solve real-world NLP problems.

We're fully back in "building new things" mode and have been thinking a lot about how to do beta testing for Ellf, our new tool for agentic NLP development, and how to make sure we're shipping updates quickly. One thing we've found from our other developer tools like spaCy and Prodigy is that it's crucial to be users of our tools as well. We need to be solving real problems and ship real projects ourselves. Otherwise, it's way too easy to work in a bubble and build things that aren't actually useful in practice. So many features and improvements were direct results of us building things and stumbling over missing functionality, unideal UX or straight-up bugs. ("Why is this so bad, who wrote this? Oh... me.") Ellf has a lot of moving parts: the coding assistant integration for project planning and NLP advice, the collaborative platform for annotation, batch jobs and agents, and the data-private cluster hosted by the user to run things and manage data. Ideally, we need to be doing and using all of it, and run as many end-to-end workflows as possible: plan a pipeline or component, create data, automate with agents, train a model, evaluate, iterate and collaborate. So here's what we're currently doing: With our beta partners, we've started scoping out practical NLP problems and are solving them at a big discount, while collaborating on development via the platform, and comparing Ellf's project advice to our solution. In the end, we get to ship real and fully independent deliverables – like a custom component, pipeline or workflow. To make sure we're getting to do many different projects, we're targeting very quick turnaround times of a few weeks and very clearly scoped deliverables. We're especially interested in use cases like information extraction and structured data, turning LLM prototypes into modular pipelines that are better, faster, cheaper and private, human-in-the-loop distillation and of course data development. So if you and your team need help with a burning NLP problem or use case and want to try out "Vibe NLP" and agentic NLP development, feel free to get in touch!