HelpMe
AI feedback inside Canvas
Instructors choose a HelpMe question inside Canvas and add it to their course. Students open the question through Canvas, submit a response, and receive course-specific AI feedback.
Shared HelpMe platform. I implemented and hardened the current LTI integration and contributed the structured feedback flow across HelpMe and Chatbot.
One question, inside the course
An instructor already has a course, a set of questions, and students working in Canvas. The integration brings HelpMe into that workflow. The instructor selects a question through a picker, and Canvas stores the resource needed to open it again.
The student follows the course link, writes a response, and asks for feedback. The embedded question keeps the course context intact. A comment and provisional score appear beside the answer, without another HelpMe login.
Built to support grading in UBC Okanagan's INDG 100, a course of about 500 students per term. That is the intended course scale. The feedback is provisional; it does not automatically become a final Canvas grade.
The instructor-to-student flow
Real interface details from the synthetic Canvas demonstration.
Bind the launch to the course
LTI 1.3 deep linking returns a signed resource carrying the selected question identifier. When that resource launches, HelpMe verifies the platform, maps the course, and checks that the question belongs to it. Those checks connect the browser flow to the course access rules.
Implementation details
I reused the ordinary LTI application session and existing course checks. A separate question-token system would have created a second authentication path to maintain. The chosen design keeps identity in the session the application already understands.
A successful LTI launch establishes application access, not Canvas API credentials. Keeping those responsibilities separate matters when deciding what a course link should be allowed to do.
Feedback that both services understand
HelpMe owns the course-specific grading rules. Chatbot returns a structured response containing a score, comment, reasons, and a review-needed field. Both services validate that contract before the result reaches the interface.
Implementation details
Invalid model output receives bounded repair attempts. If the service cannot produce valid feedback, the interface reports the problem instead of presenting a malformed answer as a grade. The review-needed field remains visible without implying that a person has already reviewed the response.
Educators control the agents
LANTERN is a separate nursing-education initiative on HelpMe. I built its configurable multi-agent architecture so educators can manage separate prompts, documents, and settings through one parent course.
Students see the parent course. Instructors manage each agent behind it, including the material used for retrieval. The educator workflow stays coherent even though the implementation uses separate agent courses.
This work connects a shared platform to specific teaching needs. My contribution is the integration and service behavior described here, alongside other builders of HelpMe and Chatbot. The demonstration uses synthetic course content and responses.