/
2026-03-16 Educators WG: AI Extensions Progress & Demo

2026-03-16 Educators WG: AI Extensions Progress & Demo

Educators WG: AI Extensions for Instructional Use Cases

Transcript

Summary

Working group discussed meeting logistics via Google Meet migration and received an extensive demonstration of EduNext’s new AI extensibility framework.

Felipe Demonstrated some early Products built on the AI Extension framework

  • User-facing chatbot that is grounded in course content.

  • Instructor-Facing Content Authoring AI to develop assessments.

  • Potential for additional products like Flashcard Generators, etc…

Introduce AI Extensibility Framework
Felipe Montoya introduced the EduNext AI extensibility framework, focusing on usability for educators, and demonstrated the Stellar Guide chatbot's ability to switch behaviors between "learning mode" and "exam mode." The framework uses Profiles, Scopes, Sessions, and Templates, with Templates being critical for configuring the assistant's behavior.

Demonstrate AI Workflows
The team saw demonstrations of the Question Builder Generator for authors and learned the framework supports Function Calling into the LMS for context retrieval, alongside an audit tool for session review. The framework is designed to function both as a suite of existing tools and as an extensible platform for building custom tools, such as the suggested Bring Your Own AI feature.

Details

  • Introduction of Felipe Montoya and AI Extensibility Discussion: John Swope introduced Felipe Montoya, the co-founder of EduNext, and praised their work on open-source initiatives like the extensions framework (00:06:33). Felipe Montoya was introduced to discuss the work EduNext has been doing on AI extensibility, focusing on aspects pertinent to educators (00:07:39).

  • Focus on AI Usability Over Technical Aspects: Felipe Montoya thanked John Swope for the introduction and stated their intention to focus the presentation on the usability of the AI extensions for educators, rather than on the technical details (00:07:39). They explained that the project seeks to break the "chicken and egg problem" of exploring AI's capacity in education by providing a technical foundation that allows for semantic and prompt capabilities to take over (00:08:39).

  • Project Alignment with Community Needs: Felipe Montoya emphasized that the AI extensions project was created to be community-first, similar to the aspects project and extensions framework, and resulted from collaboration and funding from various entities, including Axim (00:11:10). The project is available as an Open edX plugin that institutions can install and test (00:12:15).

  • Demonstration of the Stellar Guide Chatbot: Felipe Montoya demonstrated a sandbox environment called https://takoradi.edunext.cloud/ featuring a configurable chatbot named Stellar Guide within an astronomy course (00:13:28). This chatbot uses different prompts for "learning mode" and "exam mode," enabling distinct behaviors based on the course context (00:14:40).

  • Conceptual Framework of AI Extensions: The framework introduces four primary objects: Profiles (the "what" of the experience), Scopes (the "where" the assistance is present), Sessions (intermediate data storage), and Templates (reusable prompt templates) (00:17:10). Felipe Montoya highlighted that the most critical aspect is the prompt template, which configures the assistant's behavior (00:19:43).

  • Function Calling and Audit Capabilities: The LLM is allowed to call functions back into the LMS during reasoning to retrieve context, course outlines, or links (00:20:59). Felipe Montoya also showcased a debugging tool that allows users to audit sessions, revealing the system prompt, context, and the back-and-forth reasoning process of the LLM thread, which can serve as a safety net for reviewing student complaints (00:24:01).

  • Integration with Aspects and Data Collection: The AI framework is connected with aspects, automatically sending X APIs to provide data on student usage, allowing for insights into which course units utilize AI and comparing exam results between users and non-users of the AI tools (00:26:54). This is part of the foundation designed to enable educators to use LLMs effectively to improve education (00:28:18).

  • AI Workflow for Authoring Content (Question Builder): Felipe Montoya demonstrated a second type of AI workflow designed for course authors within the Studio interface, exemplified by the AI Question Builder Generator (00:28:18). This tool generates draft questions (e.g., multiple-choice) based on the unit's content, which the author can then review, edit, or discard before saving them to a content library (00:32:26).

  • Future AI Workflow Development (Badges and Flashcards): Felipe Montoya outlined plans for additional AI workflows, including one for creating badges by extracting skills and another for generating flashcards based on course content (00:36:07). The flashcard feature would allow students to assess their understanding (easy/hard) to adjust the recurrence of the cards (00:37:27).

  • Platform vs. Framework Approach for Users: John Swope asked whether potential users should view the AI Extensions primarily as a framework for building new tools or as a suite of existing tools to turn on (00:40:01). Felipe Montoya explained that it functions as both: users can leverage the provided tools or, with technical capacity, they can extend it by replacing UI components or writing custom Python "orchestrators" for sophisticated processing logic (00:40:54) (00:43:29).

  • Discussion of Bring Your Own AI Key and Customization: Michael Williams discussed their interest in "Bring Your Own AI" (BYOA) models, where students use their own keys, which reduces the cost burden on the educational institution and allows for the ingestion of student interaction data (00:46:34). Felipe Montoya responded that while BYOA key support is not currently planned, the framework is designed to allow users to build such a feature on top of the existing foundation using a custom orchestrator (00:51:46).

  • Ensuring Community Contributions and Iteration: Felipe Montoya emphasized that the framework's design prevents users from having to fork the core project, ensuring that improvements and new versions (e.g., versions 14, 15, etc.) can be easily integrated while allowing users to build and share their own solutions, like a custom orchestrator for detecting cheating or supporting mental health interventions (00:44:38) (00:53:12). John Swope requested a quick, recommended test for educators who want to try out the extensions (00:57:36).

  • Testing Options and Documentation: Felipe Montoya outlined three ways for users to engage with and test the new feature, including using the https://takoradi.edunext.cloud/ platform for easy external testing or spinning up a local instance of Tutor, since the plug-in is tutor friendly. Documentation is available at docs.openedx.org, though it is currently geared towards advanced users familiar with the admin panel rather than educators (00:58:31).

  • Engagement and Feedback: Users who have input on alternative or better technical approaches are encouraged to join the AI working group, which meets to discuss potential use cases and technical direction. John Swope noted that having the documentation available is a positive step, aligning with the Open edX ethos of documentation and integration, suggesting users can now begin exploring the topic (00:58:31).

 

Suggested next steps

  • Ildi Morris will take ownership of the next month's meeting and look for other topics.

  • Ildi Morris and Sarina Canelake will talk offline about the invite process for the next month's meeting.