Canvas AI Agent Hackathon
Overview
At the Skandalaris Hackathon 2025 (AI Track), my teammates DeYu Zhang, Mark Song, and I built Canvas AI Agent, an LLM-powered assistant designed to help students navigate course content directly within Canvas. The agent answers questions, summarizes materials, and reminds students of relevant deadlines using contextual course data. Our project earned a Runner-Up Award (Top 5 overall) and received $500 in cash prizes.
Motivation
Canvas courses often contain fragmented information spread across announcements, files, assignments, and discussion boards. As a result, students spend unnecessary time searching for information instead of focusing on learning. Our goal was to build an agent that understands course-specific context, not just raw text, and provides answers through a clean and intuitive conversational copilot interface with little learning curve, especially benefiting accessibility-constrained and younger students.
Highlights
- Integrated authorized Canvas API actions, allowing
the agent to incorporate real student-accessible course operations
directly into its responses
- Implemented retrieval-augmented generation (RAG)
using vector similarity search to accelerate discovery of relevant
Canvas materials while filtering out unrelated content
- Built a real-time frontend using WebSockets to stream incremental responses and progress updates, supporting concurrent multi-user sessions
Technical Architecture
Frontend
- React + TypeScript
- Chat-style conversational interface
- WebSocket powered real-time response streaming
Backend
- FastAPI server with asynchronous request
handling
- Azure OpenAI GPT-5 for query understanding, tool
invocation, and response generation
- OpenAI Vector Stores for semantic search across
uploaded course materials
- Canvas LMS API integration for student-accessible resources