Ellf · Your virtual NLP engineer for agentic development

Explosion, the company behind spaCy and Prodigy, launched Ellf, a virtual NLP engineer that plugs into coding assistants such as Claude Code to plan, build and evaluate task-specific NLP systems. Ellf covers project planning, distillation of large language models into smaller private components, model training and analysis, data preparation, annotation guidelines, pattern generation and QA, running on a data-private cluster the customer operates locally or in its own cloud. Explosion says coding assistants fall short on NLP in three ways — annotation, execution and knowledge — and positions Ellf as supplying those missing pieces from 10+ years of industrial NLP work.

spaCy and Prodigy Your virtual NLP engineer You’ve heard all about vibe coding – now say hello to vibe NLP . With Ellf, your team can take back control and build custom and fully private AI solutions in-house. ✳︎Claude Code › We need to extract products and features from support tickets. We started with an LLM API but it's slow and expensive, and our data needs to stay private. How can we train our own model?⏺︎Let me analyze your requirements and put together a project plan for you. Project planning Break your business problem down into actionable steps. Distillation Use LLMs to create better, smaller, faster and fully private task-specific components. Model training and analysis Train and fine-tune task-specific models, analyze the logs and diagnose problems. Data preparation and exploration Convert raw data, cluster information and iterate on label schemes for your task. Data development Collaborate on annotation and automate data labeling with agents Annotation guidelines Write comprehensive labelling manuals to scale up the data creation process. Pattern generation Determine what to extract with rule-based logic and create patterns interactively. QA and evaluation Test your system with real-world and synthetic data. Coding assistance Build custom solutions with spaCy, Prodigy and the powerful Python ecosystem. › How can I best solve this?⏺︎Let me analyze your requirements and put together a project plan for you. Use your favorite tools Ellf integrates seamlessly with your coding assistant https://ellf.ai/docs/install coding-assistants of choice and makes it proficient at solving NLP problems and executing end-to-end machine learning projects. Training experimentSTARTING Annotate entitiesSTARTED Keep your data private Run annotation, training and more, powered by a data-private cluster https://ellf.ai/docs architecture you operate locally or in your own cloud and 100% under your control. ✳︎Claude Code › /ellf-support Is this a good approach?⏺︎I've emailed your conversation logs to support@ellf.ai. Get human support If you ever feel stuck or need a second opinion, send us your logs https://ellf.ai/docs support and have them reviewed by one of our NLP experts – straight from your coding assistant. From the blog What coding assistants are missing for NLP Building software is mostly a code problem, but building an NLP system is a code plus data problem. Out of the box, coding assistants fall short in three ways – Ellf gives them the missing pieces. - Annotation: A low-friction way to get feedback from humans on real data and “schedule work on a human”. - Execution: Run long jobs and experiments on a private cluster, without building infrastructure. - Knowledge: Best practices from 10+ years of solving real-world NLP problems in industry. Read the Blog Post https://explosion.ai/blog/introducing-ellf Ellf for Prodigy users https://ellf.ai/prodigy Already using the Prodigy https://prodi.gy annotation tool? Here’s how you can scale up, automate and supercharge your annotation. Ellf for universities https://ellf.ai/universities Use Ellf to collaborate on research and experiments, and speed up and automate development and data collection. Optimize every part of your system Instead of relying on vibes, get a clear picture of what works, build a robust evaluation for every piece of your pipeline, and replace costly LLM calls with smaller, faster and private components. ✳︎Claude Code › I want to evaluate our support ticket pipeline. Can you set up a task so our team and an agent can annotate a sample, and compare the routing and extraction steps against the data?⏺︎Sure Let me set up the annotation task and an evaluation workflow for each component. Distill large modelSTARTING Get started with templates Templates package up complete end-to-end workflows so you don’t have to start from scratch. Point your coding agent to a template, adapt it to your use case and data and press play, or export your own projects to share with your team. ✳︎Claude Code › /ellf-template Set up this project using the sample data: https://github.com/explosion/ellf-templates/tree/main/llm-to-jev⏺︎I've read the template's README. Let me set up the project, tasks and evaluation workflow. Build custom NLP with Claude Code and agents How do you get from an LLM prototype to a fast, cheap and fully private NLP component you can run in production? In this video, Ines shows you how Ellf helps coding agents like Claude Code build a custom NLP model to analyze incoming support tickets. Watch on YouTube https://www.youtube.com/watch?v=WvlUl5c9-zM&list=PLLU8Ghv6hhVA&index=2