Vibe NLP for Applied NLP

One of the hardest parts of applied NLP has always been breaking down complex business problems into machine learning components. It's so hard because it requires domain expertise and reasoning about the specific use case, and it's the one thing technology couldn't fix. But what if we could take some of the learnings from AI-powered coding assistants and apply them to solving real-world NLP problems? In this talk, I'll show how we've built powerful assistants and MCP servers to help developers solve NLP tasks using open-source software, and create practical and modular solutions that are small, fast and fully data-private.

At the core of it is an often overlooked idea: using LLMs to build systems instead of as systems. AI-powered coding assistants have transformed the way we build software – and they can be even more impactful for AI development itself and bridge the experience gap that's often holding teams back and causing projects to fail.

Transcript

1. Ines Montani Explosion
2. Open-source library for industrial-strength natural language processing spacy.io 570m+ downloads
3. Open-source library for industrial-strength natural language processing spacy.io LLMs are really good at spaCy code! 570m+ downloads
4. users companies Modern scriptable annotation tool for machine learning developers prodigy.ai 12k+ 1000+
5. users companies fully scriptable in Python Alex Smith Developer Kim Miller Analyst GPT-5 API Modern scriptable annotation tool for machine learning developers prodigy.ai 12k+ 1000+
6. of coding assistants
7. help developer implement code for the given tools of coding assistants
8. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants
9. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants solve a business problem
10. help developer implement code for the given tools help developer pick the right tools and implement code of coding assistants solve a business problem “I need to analyze these company reports and create a table of the total spending on di ff erent types of IT services over time.” 2025.pdf 2024.pdf 2023.pdf
11. reverse these strings
12. reverse these strings write a script to reverse strings
13. reverse these strings write a script to reverse strings prompt program
14. list all company names in the text write a script to extract company names from text
15. as the system
16. as the system to build
17. sort these documents into custom categories create data and train a classifier for custom categories
18. Why don’t we just build ?
19. Why don’t we just build ? lots of steps
20. Why don’t we just build ? lots of steps operational complexity
21. Why don’t we just build ? lots of steps operational complexity expertise
22. Why don’t we just build ? lots of steps operational complexity expertise coding agents are really good at software 1.0 Software 1.0
23. Why don’t we just build ? lots of steps operational complexity expertise coding agents are really good at software 1.0 Software 1.0 kind of
24. Why don’t we just build ? lots of steps operational complexity expertise coding agents are really good at software 1.0 Software 1.0 Software 2.0 code + data kind of
25. Why don’t we just build ? lots of steps operational complexity expertise coding agents are really good at software 1.0 Software 1.0 Software 2.0 code + data agentic kind of
26. ✳ Claude Code beta.ellf.ai
27. ✳ Claude Code beta.ellf.ai ❯ We need to extract product mentions and pricing from company press releases. We started with an LLM inference API but it's too slow and expensive, plus we can't send confidential data to an external API. How can we train our own model?
28. ✳ Claude Code beta.ellf.ai ❯ We need to extract product mentions and pricing from company press releases. We started with an LLM inference API but it's too slow and expensive, plus we can't send confidential data to an external API. How can we train our own model? ⏺ Yes, and this is exactly the right call. What you're describing — structured extraction from a known document type — is precisely where a supervised model beats an LLM on every dimension: speed, cost, privacy, and accuracy once you have enough examples. The LLM you've been using is essentially doing a job that a small NER model will do 10–100× faster, for low marginal cost, entirely on-premise. My rough pipeline sketch for your case:
29. ✳ Claude Code beta.ellf.ai distilled expertise from 10+ years of solving NLP problems ❯ We need to extract product mentions and pricing from company press releases. We started with an LLM inference API but it's too slow and expensive, plus we can't send confidential data to an external API. How can we train our own model? ⏺ Yes, and this is exactly the right call. What you're describing — structured extraction from a known document type — is precisely where a supervised model beats an LLM on every dimension: speed, cost, privacy, and accuracy once you have enough examples. The LLM you've been using is essentially doing a job that a small NER model will do 10–100× faster, for low marginal cost, entirely on-premise. My rough pipeline sketch for your case:
30. ✳ Claude Code beta.ellf.ai distilled expertise from 10+ years of solving NLP problems PROJECT_PLAN.md ❯ We need to extract product mentions and pricing from company press releases. We started with an LLM inference API but it's too slow and expensive, plus we can't send confidential data to an external API. How can we train our own model? ⏺ Yes, and this is exactly the right call. What you're describing — structured extraction from a known document type — is precisely where a supervised model beats an LLM on every dimension: speed, cost, privacy, and accuracy once you have enough examples. The LLM you've been using is essentially doing a job that a small NER model will do 10–100× faster, for low marginal cost, entirely on-premise. My rough pipeline sketch for your case:
31. ✳ Claude Code ❯ / beta.ellf.ai/docs/modules
32. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem beta.ellf.ai/docs/modules
33. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data beta.ellf.ai/docs/modules
34. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data /ellf-prodigy Implement workflows for the Prodigy annotation tool beta.ellf.ai/docs/modules
35. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data /ellf-prodigy Implement workflows for the Prodigy annotation tool /ellf-patterns Interactively develop patterns and rule-based logic beta.ellf.ai/docs/modules
36. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data /ellf-prodigy Implement workflows for the Prodigy annotation tool /ellf-patterns Interactively develop patterns and rule-based logic /ellf-train Train and export custom models and components beta.ellf.ai/docs/modules
37. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data /ellf-prodigy Implement workflows for the Prodigy annotation tool /ellf-patterns Interactively develop patterns and rule-based logic /ellf-train Train and export custom models and components /ellf-ask Ask specific questions and query the knowledge base beta.ellf.ai/docs/modules
38. ✳ Claude Code ❯ / /ellf-project Plan the project and break down business problem /ellf-annotate Define labels, components and annotate data /ellf-prodigy Implement workflows for the Prodigy annotation tool /ellf-patterns Interactively develop patterns and rule-based logic /ellf-train Train and export custom models and components /ellf-ask Ask specific questions and query the knowledge base beta.ellf.ai/docs/modules will be open- sourced soon! skills
39. We need to run lots of stu ff !
40. training We need to run lots of stu ff !
41. training We need to run lots of stu ff ! optimization
42. training We need to run lots of stu ff ! analysis optimization
43. training evaluation We need to run lots of stu ff ! analysis optimization
44. training data transformation evaluation We need to run lots of stu ff ! analysis optimization
45. training data transformation evaluation We need to run lots of stu ff ! analysis testing optimization
46. training data transformation evaluation We need to run lots of stu ff ! analysis testing optimization annotation tasks auto-annotation agents
47. training data transformation evaluation various scripts We need to run lots of stu ff ! analysis testing optimization annotation tasks auto-annotation agents
48. training data transformation evaluation various scripts We need to run lots of stu ff ! … analysis testing optimization annotation tasks auto-annotation agents
49. training data transformation evaluation various scripts We need to run lots of stu ff ! … analysis testing optimization annotation tasks auto-annotation agents
50. training data transformation evaluation various scripts We need to run lots of stu ff ! … analysis testing GPU optimization annotation tasks auto-annotation agents
51. training data transformation evaluation various scripts We need to run lots of stu ff ! … analysis testing often sensitive data GPU optimization annotation tasks auto-annotation agents
52. training data transformation evaluation various scripts We need to run lots of stu ff ! … analysis testing often sensitive data GPU fully custom optimization annotation tasks auto-annotation agents
53. Kubernetes cluster beta.ellf.ai
54. hosted by user local or cloud Kubernetes cluster beta.ellf.ai
55. hosted by user local or cloud Kubernetes cluster tasks beta.ellf.ai
56. hosted by user local or cloud Kubernetes cluster tasks beta.ellf.ai
57. hosted by user local or cloud Kubernetes cluster tasks actions beta.ellf.ai
58. hosted by user local or cloud Kubernetes cluster tasks actions beta.ellf.ai
59. hosted by user local or cloud Kubernetes cluster tasks actions agents beta.ellf.ai
60. hosted by user local or cloud Kubernetes cluster tasks actions agents beta.ellf.ai
61. beta.ellf.ai
62. beta.ellf.ai humans
63. beta.ellf.ai humans agents
64. Human + coding agent collaboration beta.ellf.ai
65. Human + coding agent collaboration @task_recipe( title="My Recipe", field_props={ "input": TextProps(title="Input data"), "labels": ListProps(title="Labels"), "dataset": props.dataset_choice, } ) def my_recipe(input: Input, labels: List[str], dataset: Dataset): ... Recipe beta.ellf.ai
66. Human + coding agent collaboration @task_recipe( title="My Recipe", field_props={ "input": TextProps(title="Input data"), "labels": ListProps(title="Labels"), "dataset": props.dataset_choice, } ) def my_recipe(input: Input, labels: List[str], dataset: Dataset): ... Recipe $ ellf tasks create my_recipe -- input "Corpus" - - labels PRODUCT,ORG - - dataset products CLI beta.ellf.ai
67. Human + coding agent collaboration @task_recipe( title="My Recipe", field_props={ "input": TextProps(title="Input data"), "labels": ListProps(title="Labels"), "dataset": props.dataset_choice, } ) def my_recipe(input: Input, labels: List[str], dataset: Dataset): ... Recipe $ ellf tasks create my_recipe -- input "Corpus" - - labels PRODUCT,ORG - - dataset products CLI Input data input Corpus Labels labels PRODUCT ORG Dataset dataset products UI Start task beta.ellf.ai
68. Human + coding agent collaboration @task_recipe( title="My Recipe", field_props={ "input": TextProps(title="Input data"), "labels": ListProps(title="Labels"), "dataset": props.dataset_choice, } ) def my_recipe(input: Input, labels: List[str], dataset: Dataset): ... Recipe $ ellf tasks create my_recipe -- input "Corpus" - - labels PRODUCT,ORG - - dataset products CLI scripting the SaaS application! Input data input Corpus Labels labels PRODUCT ORG Dataset dataset products UI Start task beta.ellf.ai
69. beta.ellf.ai App and developer API need to speak the same language!
70. Chat with Ellf beta.ellf.ai App and developer API need to speak the same language!
71. Chat with Ellf beta.ellf.ai App and developer API need to speak the same language! machine human
72. Chat with Ellf ✳ Claude Code ❯ /ellf-todo ⏺ Context summary: Phase 1 of an e-commerce aspect extraction project. The task is to write loader.py for the internal data warehouse API that segments sentences and yields {id, text, date} dictionaries. ✔ Successfully updated coding request 523060d8-b2db-4fd5-b541-5ee119007501 beta.ellf.ai App and developer API need to speak the same language! machine human
73. Software 2.0 agentic We’ll still need...
74. Software 2.0 agentic We’ll still need... software
75. Software 2.0 agentic We’ll still need... software ML systems
76. Software 2.0 agentic We’ll still need... software ML systems classifiers
77. Software 2.0 agentic We’ll still need... software ML systems interfaces for humans classifiers
78. Use LLMs to build the system, not as the system.
79. Use LLMs to build the system, not as the system. Apps, developer APIs and agents need to speak the same language.
80. Use LLMs to build the system, not as the system. Apps, developer APIs and agents need to speak the same language. Software 2.0 is a good idea and we can finally make it easy!
81. explosion.ai spacy.io prodigy.ai beta.ellf.ai @inesmontani.bsky.social @ines@sigmoid.social Explosion spaCy Prodigy Ellf Bluesky Mastodon LinkedIn
82. explosion.ai spacy.io prodigy.ai beta.ellf.ai @inesmontani.bsky.social @ines@sigmoid.social Explosion spaCy Prodigy Ellf Bluesky Mastodon more details and waitlist LinkedIn