TA Reply Copilot

Course information is scattered across Brightspace announcements, the syllabus, PDFs, and rubrics. TAs face repeated student questions about deadlines, late policies, project scope, and presentation rules. Generic chatbots may hallucinate policies, which is risky in an academic context.

A lightweight assistant that ingests course materials from Brightspace exports (announcements, content, PDFs), and given a student email produces a suggested reply (ready to send), cited evidence (source + date/page + snippet), and a list of missing info to ask the student if the question is underspecified.

• Paste the student email into the system
• It retrieves relevant course text via RAG
• Generates a reply draft with citations
• The TA edits and sends
• Goal: cut TA email time while keeping replies policy-correct and verifiable

• Retrieval-Augmented Generation — over course documents
• Citation generation — every claim links to a source + page/date
• Underspecified-question detection — surfaces missing info to clarify

• Data: Brightspace exports (announcements, content) + PDFs
• Parsing: HTML-to-text + PyMuPDF / pdfplumber
• Retrieval: Chroma / FAISS (+ optional BM25)
• Frontend: React or minimal web frontend
• Optional: Chrome extension stretch goal
• Vibe coding: Cursor / Claude Code / ChatGPT